Compiler-assisted checkpointing of message-passing …gac.udc.es/tesis/GabrielRodriguezAlvarez.pdf · 2011-12-13 · Ejemplos de estas herramientas son Calypso [8] y las extensiones - [PDF Document] (2024)

Compiler-assisted checkpointing ofmessage-passing applications in

heterogeneous environments

Gabriel Rodríguez Álvarez

Department of Electronics and Systems

University of A Coruña, Spain

Department of Electronics and Systems

University of A Coruña, Spain

Phd Thesis

Compiler-assisted checkpointing of

message-passing applications in

heterogeneous environments

Gabriel Rodríguez Álvarez

Diciembre de 2008

PhD Advisors:María J. Martín Santamaríaand Patricia González Gómez

Dra. María J. Martín SantamaríaProfesora Titular de UniversidadDpto. de Electrónica e SistemasUniversidade da Coruña

Dra. Patricia González GómezProfesora Titular de UniversidadDpto. de Electrónica e SistemasUniversidade da Coruña

CERTIFICAN

Que la memoria titulada Compiler-assisted checkpointing of message-passing appli-

cations in heterogeneous environments ha sido realizada por D. Gabriel RodríguezÁlvarez bajo nuestra dirección en el Departamento de Electrónica e Sistemas de laUniversidade da Coruña y concluye la Tesis Doctoral que presenta para optar algrado de Doctor en Informática.

A Coruña, 15 de Diciembre de 2008

Fdo.: María J. Martín SantamaríaCodirectora de la Tesis Doctoral

Fdo.: Patricia González GómezCodirectora de la Tesis Doctoral

Fdo.: Luís Castedo RibasDirector del Dpto. de Electrónica e Sistemas

Resumen de la tesis

Introducción

La evolución de la computación paralela hacia infrastructuras cluster y Grid hacreado nuevas necesidades de tolerancia a fallos. A medida que las máquinas parale-las incrementan su número de procesadores, también se incrementa la tasa de fallosdel sistema global. Esto no supone un problema mientras el tiempo medio que unaaplicación tarda en terminar su ejecución permanezca por debajo del tiempo mediohasta fallo (MTTF ) del hardware, pero esto no siempre se cumple para aplicacionescon ejecuciones largas. En estas circunstancias, los usuarios y programadores necesi-tan disponer de mecanismos que les permitan garantizar que no toda la computaciónrealizada se pierde al ocurrir un fallo.

El checkpointing se ha convertido en una técnica ampliamente utilizada para laobtención de tolerancia a fallos. Consiste en el almacenamiento periódico del estadode la computación, de forma que la ejecución pueda ser restaurada a partir de dichoestado en caso de fallo. Se han propuesto diversas soluciones y técnicas [22] paralograr este propósito, cada una con sus propias ventajas e inconvenientes.

Las técnicas de checkpointing aparecieron originalmente como servicios propor-cionados por el sistema operativo (SO), centradas en la recuperación de aplicacionessecuenciales. Ejemplos de este tipo de herramientas son KeyKOS [36], que realizabacheckpointing a nivel del sistema, almacenando el estado del operativo completo,y Sentry [56], una implementación basada en UNIX que realizaba checkpointing yregistro de eventos no deterministas para aplicaciones secuenciales. Sprite [45] secentraba en la gestión de la migración de procesos para el checkpoint y reinicio deprocesos en ejecución en computadores de memoria compartida. Su objetivo era ba-lancear la carga de una red de estaciones de trabajo a través de la migración deprocesos que ejecutaban una única aplicación de memoria compartida a máquinas

v

vi

con baja carga. Al estar implementados en el operativo, estas herramientas erancompletamente ad-hoc, con falta de énfasis en la portabilidad. Su aproximación a laeciencia consistía en obtener un buen rendimiento de E/S para el almacenamientodel estado de la computación, en lugar de reducir la cantidad de datos a almacenar.

La primera desventaja obvia de las implementaciones basadas en el operativo esla dependencia existente entre la tolerancia a fallos y el sistema operativo concretoen uso. Las herramientas de checkpointing comunes en los operativos hasta los 90,comenzaron a desparecer en entornos populares como UNIX, SunOS o AIX. La ne-cesidad de soluciones exibles capaces de operar en entornos diferentes motivaronla aparición de las soluciones a nivel de aplicación (opuestas a las soluciones a nivelde sistema). En estas herramientas, la tolerancia a fallos se consigue a través dela compilación de la aplicación junto con el código de checkpointing ubicado habi-tualmente en una librería aparte. Las herramientas de checkpointing en este períodoeran aún transparentes, almacenando el estado de la aplicación al completo. Al noimplementarse dentro del kernel del operativo debían enfrentarse a problemas im-portantes al manipular estado dependiente del SO. Ejemplos son la recuperación delos identicadores de procesos o de la tabla de cheros abiertos. Además, debíandisponer de mecanismos de recuperación de la pila o el montículo. Estos problemashacían que el código de las herramientas fuese aún muy dependiente de los serviciosproporcionados por el operativo. Normalmente, esto forzaba a los desarrolladoresa restringir el tipo de servicios que una aplicación podía usar. Ejemplos de herra-mientas transparentes a nivel de aplicación son Libckpt [49], CATCH GCC [38] (unaversión modicada del compilador GNU C), y la herramienta uniprocesador incluídacon el sistema Condor [41].

A mitad de los 90 comenzaron a aparecer soluciones no transparentes que inten-taban aplicar checkpointing a plataformas de computación distribuida. Su desven-taja fundamental era la falta de una interfaz estándar para la comunicación entreprocesos. Ejemplos de estas herramientas son Calypso [8] y las extensiones a Do-me (Distributed Object Migration Enviroment) [10]. En ambos casos el lenguaje deprogramación objetivo era una extensión de C++ con construcciones paralelas noestándar.

La adopción de MPI como el estándar de-facto para programación paralela hamotivado la aparición de herramientas de checkpointing basadas en esta interfaz enlos últimos años. Al principio, estas herramientas usaban la aproximación transpa-rente a nivel de aplicación, compartiendo las desventajas de sus primas para aplica-ciones secuenciales: falta de portabilidad de los datos y restricción de los entornos

vii

soportados, que aquí se reere a la implementación concreta de MPI. De hecho, loscheckpointers en esta categoría se implementan generalmente mediante la modi-cación de una librería MPI previamente existente. Ejemplos de este tipo de check-pointers son MPICH-GF [71] y MPICH-V2 [12], ambas implementadas como driversMPICH, obligando por tanto a todas las máquinas a disponer de esta implementa-ción concreta de MPI.

Las nuevas tendencias de computación, tales como el uso de grandes clusters

heterogéneos y sistemas Grid, presentan nuevas restricciones para las técnicas decheckpointing. La heterogeneidad hace imposible aplicar soluciones tradicionales dealmacenamiento del estado de la computación, que usan estrategias no portablespara la recuperación de estructuras como la pila, el montículo, o el estado de lascomunicaciones de las aplicaciones.

Así, las nuevas técnicas de checkpointing deben proporcionar estrategias para larecuperación portable del estado de la ejecución, donde ésta pueda ser restaurada enun amplio rango de máquinas, ya sean éstas incompatibles a nivel binario o usen ver-siones incompatibles de utilidades software, tales como interfaces de comunicacionesentre procesos.

Este trabajo presenta CPPC (ComPiler for Portable Checkpointing), una herra-mienta de checkpointing centrada en la inserción automática de tolerancia a fallosen aplicaciones de paso de mensajes. Ha sido diseñada para permitir el reiniciode las ejecuciones en diferentes arquitecturas y/o sistemas operativos, permitiendotambién el checkpointing en sistemas heterogéneos. Utiliza código y protocolos por-tables, y genera cheros de estado portables al mismo tiempo que evita el uso desoluciones tradicionales, como la coordinación de procesos o el registro de mensajes(message-logging) que están ligadas a falta de escalabilidad.

Metodología de trabajo

El primer paso en la elaboración de la tesis ha sido la revisión de la bibliografíaexistente en el ámbito de la tolerancia a fallos de aplicaciones paralelas. Estableci-das las ventajas y general predominancia de las soluciones basadas en checkpointing,se ha abordado el diseño partiendo de aplicaciones extremadamente sencillas, cons-truyendo soluciones ad-hoc y continuando con ejemplos cada vez más complejos ytratando de abstraer lo común a todas ellas. A partir de estas experiencias, se haconstruido una librería que proporcione apoyo a las tareas comunes necesarias, que

viii

fue posteriormente probada mediante utilización manual para la inserción de tole-rancia a fallos en diversas aplicaciones cientícas de complejidad superior a las yaestudiadas. En caso de detectar alguna falla, esta fase ha servido para mejorar lalibrería, realizando las mejoras necesarias que podían preverse en la etapa de análisisinicial.

Una vez disponible la librería funcional utilizada para la obtención manual detolerancia a fallos, se ha procedido a trabajar en la construcción de la herramientade transformación, haciendo uso del conocimiento adquirido mediante la transfor-mación manual de las aplicaciones, generalizando las técnicas utilizadas y realizandoimplementaciones automáticas de dichas transformaciones.

En todo momento, se ha procurado, con vistas a cumplir el objetivo de usabilidaden entornos heterogéneos, que las soluciones implementadas tengan un diseño tal quepermita su utilización en diferentes entornos de trabajo, bien mediante técnicas deingeniería de software especícas para dicho n o mediante el uso de herramientasde programación que permitan tal característica de forma nativa.

Una vez implementado el grueso de la infraestructura, se ha procedido a realizaruna metódica evaluación experimental de su comportamiento.

Contribuciones

Este trabajo presenta técnicas novedosas para el checkpointing de aplicacionesparalelas, dirigidas a la obtención de las siguientes propiedades relevantes de tole-rancia a fallos:

1. Independencia del SO: las estrategias de checkpointing deben ser compatiblescon un amplio espectro de sistemas operativos. Esto implica el disponer de,al menos, una estructura modular básica que permita la sustitución de ciertassecciones críticas de código (p. ej. acceso al sistema de cheros) dependiendodel SO subyacente.

2. Soporte de aplicaciones paralelas con independencia del protocolo de comu-nicaciones: la herramienta de checkpointing no debería presuponer el uso deuna interfaz de comunicaciones determinada. Los Grid, por ejemplo, incluyenmáquinas pertenecientes a diferentes entidades independientes que no puedenser obligadas a incluir una versión determinada de la interfaz MPI. Incluso

ix

reconociendo el rol de MPI como estándar de-facto para el paso de mensajes,la técnica de checkpointing no puede estar ligada a nivel teórico con MPI sidebe proporcionar una solución portable y reutilizable.

3. Reducción del tamaño de los cheros de estado: la herramienta debe optimi-zar la cantidad de datos almacenados, evitando incluir datos no necesariospara el reinicio de la aplicación. Esto mejora el rendimiento, que depende engran medida del tamaño de los cheros. También mejora el rendimiento de lamigración de los procesos, en caso de ser necesaria.

4. Recuperación de datos portable: el estado de la aplicación debe ser visto comouna estructura que contiene diversos tipos de datos. La herramienta de che-

ckpointing debe ser capaz de recuperar todos estos datos de forma portable.Esto incluye la recuperación de estado opaco, como comunicadores MPI, asícomo de estado dependiente del operativo, como la tabla de cheros abiertoso la pila.

Con respecto a la independencia del SO, se plantea la construcción de una herra-mienta con una elevada carga de modularidad, que posibilite el reemplazo dinámicode la implementación de diversas utilidades como puede ser el acceso al sistema decheros o las comunicaciones entre procesos. En este aspecto cobra importancia elprotocolo de reinicio propuesto, basado en la re-ejecución selectiva de código para larecuperación de las partes críticas del estado que no pueden ser simplemente alma-cenadas mediante el uso de formatos portables. Para la portabilidad del compilador,se emplean técnicas basadas en el suministro de datos referentes a las semánticas delas diversas funciones utilizadas por la aplicación a instrumentar (catálogos semánti-cos). El suministro de estos datos se realiza mediante cheros XML que documentanlibrerías de funciones ampliamente utilizadas (p. ej. funciones POSIX ) proporcio-nados junto con la herramienta, de forma que el usuario no sea responsable de sucreación.

El soporte de aplicaciones paralelas con independencia del protocolo de comu-nicaciones utilizado se posibilita, además de por el uso de los catálogos semánticosanteriormente mencionados, mediante el uso de la técnica de coordinación estáticade los procesos en tiempo de compilación. La creación de los cheros de estado enpuntos seguros proporciona, además, un alto nivel de escalabilidad a la herramienta.El compilador incluye análisis de ujo de comunicaciones y de complejidad de códigoque posibilitan la automatización del proceso de inserción de checkpoints en dichospuntos seguros.

x

La reducción del tamaño de los cheros, y por tanto la operación eciente, seconsigue mediante el almacenamiento de los datos a nivel de variable, esto es, losdatos almacenados en los cheros de estado son variables de usuario y metadatosque posibilitan la recuperación de éstos durante el proceso de re-ejecución selectivallevado a cabo en el reinicio de las aplicaciones. El compilador incluye análisis devariables vivas que automatizan la selección de las variables de usuario a almacenar.

Para la recuperación portable de todos los datos de la aplicación se introduce,aparte del protocolo de reinicio mediante re-ejecución selectiva, un formato jerárqui-co de los cheros que permite la recuperación de los datos en el ámbito de ejecuciónadecuado. El formato permite también almacenar los determinantes de los eventosno deterministas (en este caso modelados como la ejecución de rutinas no portables)de forma que dichos eventos puedan ser re-ejecutados durante el reinicio en idénti-cas condiciones, restaurando así el estado de forma semánticamente apropiada y conindependencia de la implementación concreta de las rutinas no portables objeto dela re-ejecución.

Este trabajo presenta también una evaluación experimental de la herramientanovedosa, dado que se ha realizado en un entorno de computación de altas pres-taciones público (el Finis Terrae del Centro de Supercomputación de Galicia), adiferencia de muchas otras herramientas que realizan su evaluación en entornos ce-rrados y altamente controlados. Para ello, ha sido necesaria la realización de unelevado volumen de experimentos junto con el uso de técnicas de análisis estadísticode los datos arrojados por los mismos. Las pruebas se han realizado sobre un grannúmero de códigos públicamente disponibles, y se ha proporcionado también grancantidad de información sobre la instrumentación introducida en cada código, deforma que sea posible la repetición y el contraste de los experimentos.

Conclusiones

CPPC es una herramienta de checkpointing transparente para aplicaciones depaso de mensajes con ejecuciones largas. Está formada por una librería que contie-ne rutinas de checkpointing, junto con un compilador que automatiza el uso de lalibrería. Las aspectos más destacables de esta infraestructura son:

El protocolo de coordinación en tiempo de compilación: la consistencia globalde los diferentes cheros de estado locales se consigue a través de la ubicación

xi

en puntos seguros de los checkpoints. De este modo, las operaciones relaciona-das con la consistencia se transeren de la ejecución normal de las aplicacionesa la compilación y el reinicio de las mismas. Durante la compilación, se detec-tan los puntos seguros y se seleccionan algunos de ellos para la introducciónde checkpoints. Durante el reinicio los procesos involucrados en la ejecuciónde la aplicación se comunican entre sí y deciden desde cuál de los checkpointsdebe realizarse el dicho reinicio. La sincronización de procesos utilizada porlas aproximaciones tradicionales se realiza durante el reinicio, mejorando asíla escalabilidad de la solución. Además, se mejora la eciencia, dado que tantola compilación como el reinicio son operaciones mucho menos frecuentes quela creación de los cheros de estado.

La reducción del tamaño de los cheros de estado: CPPC utiliza una aproxi-mación a nivel de variable, almacenando sólo aquellas variables que son nece-sarias durante el reinicio. Al restringir la cantidad de datos almacenados, sedisminuye el tamaño de los cheros de estado y se reduce la sobrecarga delcheckpointing. Esto mejora también el rendimiento de las transferencias de losmismos a través de la red, si fuera necesario.

La recuperación portable del estado de la aplicación: la recuperación del es-tado se realiza de forma portable mediante el formato jerárquico utilizadopara el volcado de los datos y el proceso de re-ejecución selectiva de códigono-portable que tiene lugar durante el reinicio. Esta re-ejecución proporcionatambién mecanismos para la recuperación del estado de librerías externas enuso por la aplicación. La portabilidad es una propiedad interesante debido ala inherente heterogeneidad de las arquitecturas en boga para la computaciónde altas prestaciones, como el Grid. CPPC-G [55] es un proyecto en desarrollobasado en la creación de una arquitectura de servicios web para dar soporte ala ejecución de aplicaciones CPPC en entornos Grid.

El diseño modular de la librería: la librería CPPC está implementada en C++,utilizando un diseño altamente modular que permite la conguración exiblede todas sus funcionalidades. Permite congurar dinámicamente el formatode los cheros de estado, así como activar varias características opcionalescomo la compresión de los mismos, importante cuando se trabaja sobre re-des lentas o con espacio en disco limitado, o comprabaciones de integridad.El diseño modular permite también la utilización de la herramienta en entor-nos donde no se dispone de derechos de administración, lo que es común al

xii

trabajar con computadores de altas prestaciones. El uso del patrón modelo-vista-controlador permite las llamadas a la librería desde diversos lenguajesde programación, mediante la implementación de una delgada capa de softwa-re que adapte las peticiones de la aplicación a la interfaz interna del núcleoCPPC.

El checkpointing completamente transparente: el compilador CPPC conviertea la librería CPPC en una herramienta de checkpointing totalmente transpa-rente, pues automatiza todos los análisis y transformaciones necesarios parael uso de la misma. El compilador incluye también una novedosa técnica pa-ra la identicación automática de lazos intensivos en computación basada enel análisis de métricas de complejidad de código típicas de la ingeniería delsoftware.

CPPC ha sido evaluado experimentalmente demostrando su usabilidad, escala-bilidad, eciencia y portabilidad. Es capaz de reiniciar aplicaciones en diferentesarquitecturas y/o máquinas utilizando diferentes compiladores C/Fortran e imple-mentaciones MPI, todo ello usando el mismo conjunto de cheros de estado. Elrendimiento de la herramienta se ha estimado usando aproximaciones estadísticassobre los resultados obtenidos de ejecuciones en una infraestructura de computaciónpública (el supercomputador Finis Terrae del CESGA).

Hasta donde alcanza nuestro conocimiento, CPPC es el único checkpointer deaplicaciones de paso de mensajes disponible en el dominio público. CPPC es unproyecto de código abierto, disponible en http://cppc.des.udc.es bajo licenciaGPL.

El desarrollo de esta tesis ha dado lugar a las siguientes publicaciones:

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. Controla-dor/preCompilador de Checkpoints Portables. En Actas de las XV Jornadas

de Paralelismo, pp. 253258, Almería, Septiembre de 2004.

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. On desig-ning portable checkpointing tools for large-scale parallel applications. En Pro-

ceedings of the 2nd International Conference on Computational Science and

Engineering (ICCSE'05), pp. 191203. Estambul (Turquía), Junio de 2005.

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. PortableCheckpointing of MPI applications. En Proceedings of the 12th Workshop on

xiii

Compilers for Parallel Computers (CPC'06), pp. 396410, A Coruña, Enerode 2006.

G. Rodríguez, M.J. Martín, P. González y J. Touriño. Controller/Precompilerfor Portable Checkpointing. En IEICE Transactions on Information and Sys-

tems, Vol. E89-D, No. 2, pp. 408417, Febrero de 2006.

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. CPPC:Una herramienta portable para el checkpointing de aplicaciones paralelas. enBoletín de la red nacional de I+D, RedIRIS, No. 80, pp. 5761, Abril de 2007.

D. Díaz, X.C. Pardo, M.J. Martín, P. González y G. Rodríguez. CPPC-G:Fault-Tolerant Parallel Applications on the Grid. En Proceedings of the 1st

Iberian Grid Infrastructure Conference (IBERGRID'07), pp. 230241, Santia-go de Compostela, Mayo de 2007.

G. Rodríguez, P. González, M.J. Martín y J. Touriño. Enhancing Fault-Tolerance of Large-Scale MPI Scientic Applications. En Proceedings of the

9th International Conference on Parallel Computing Technologies (PaCT'07),volumen 4671 de Lecture Notes in Computer Science, pp. 153161, Pereslavl-Zalessky (Rusia), Septiembre de 2007.

D. Díaz, X.C. Pardo, M.J. Martín, P. González y G. Rodríguez. CPPC-G: FaultTolerant Parallel Applications on the Grid. En 3rd Workshop on Large ScaleComputations on Grids (LaSCoG'07). Lecture Notes in Computer Science,volumen 4967, pp. 852859, Mayo de 2008. ISBN 978-3-540-68105-2.

G. Rodríguez, X.C. Pardo, M.J. Martín, P. González, D. Díaz. A Fault Tole-rance Solution for Sequential and MPI Applications on the Grid. En Scalable

Computing: Practice and Experience, Vol. 9, No. 2, pp. 101109, Junio de 2008.ISSN 1895-1767.

G. Rodríguez, M.J. Martín, P. González, J. Touriño, R. Doallo. CPPC: ACompiler-Assisted Tool for Portable Checkpointing of Message-Passing Appli-cations. En Proceedings of the 1st International Workshop on Scalable Tools for

High-End Computing (STHEC'08), held in conjunction with the 22nd ACM

International Conference on Supercomputing (ICS'08), pp. 112, Kos (Grecia),Junio de 2008.

xiv

Trabajo Futuro

Hay dos aspectos fundamentales en los que mejorar la herramienta CPPC. Enprimer lugar, la coordinación espacial de procesos tiene una desventaja importante:la necesidad de especicar la frecuencia de checkpointing como un parámetro depen-diente del número de llamadas a la función de checkpointing, en lugar de en funcióndel tiempo. Actualmente estamos trabajando en un protocolo ligero y no coordinadode comunicaciones que permite variar dinámicamente la frecuencia espacial en fun-ción de una frecuencia temporal especicada por el usuario. Este protocolo se basaen la comunicación no bloqueante entre todos los procesos de información relativaa la velocidad a la que cada uno de ellos ejecuta el código para, posteriormente,adoptar una frecuencia espacial que garantice que el proceso más lento crea cherosde estado con una frecuencia similar a la especicada.

La segunda mejora está relacionada con los análisis de complejidad llevados acabo para la inserción de checkpoints en el código. El algoritmo actualmente usado enel compilador CPPC puede ser mejorado a través del uso de métricas más complejas,de forma que sea posible obtener resultados más próximos a los óptimos.

Abstract

With the evolution of high performance computing towards heterogeneous, mas-sively parallel systems, parallel applications have developed new checkpoint andrestart necessities. Whether due to a failure in the execution or to a migrationof the processes to dierent machines, checkpointing tools must be able to oper-ate in heterogeneous environments. However, some of the data manipulated by aparallel application are not truly portable. Examples of these include opaque state(e.g. data structures for communications support) or diversity of interfaces for asingle feature (e.g. communications, I/O). Directly manipulating the underlyingad-hoc representations renders checkpointing tools incapable of working on dier-ent environments. Portable checkpointers usually work around portability issuesat the cost of transparency: the user must provide information such as what dataneeds to be stored, where to store it, or where to checkpoint. CPPC (ComPiler forPortable Checkpointing) is a checkpointing tool designed to feature both portabilityand transparency, while preserving the scalability of the executed applications. Itis made up of a library and a compiler. The CPPC library contains routines forvariable level checkpointing, using portable code and protocols. The CPPC com-piler achieves transparency by relieving the user from time-consuming tasks, suchas performing code analyses and adding instrumentation code.

Index terms Fault tolerance, checkpointing, parallel programming, message-passing, MPI, compiler support.

Acknowledgements

Many people have contributed to the work presented in this thesis. My Ph.D.advisors, María and Patricia, have made it possible with their constant support andhard work, as well as their sheer patience. Although not an advisor, Juan has alsogreatly contributed to the development of this work through both his eorts and hisinvaluable sense of criticism. I also want to extend my thanks to the people at theComputer Architecture Group, particularly to its head Ramón Doallo, and at theDepartment of Electronics and Systems.

On a personal level, this work would not have been the same without the solidanchorage point my family provides. And certanly not the same without María,who should be thanked for all the things that do not t on this page. Also on apersonal level albeit admittedly a geeky one Marcos and José have a habit ofmaking things always brighter and sharper through their ideas, criticism, and alsotheir moaning and whining (strictly when necessary).

I also want to acknowledge the following persons and institutions: CESGA (Gali-cian Supercomputing Center) for providing access to their computing resources, par-ticularly the Finis Terrae supercomputer, and specially to José Carlos Mouriño forhelping me out in my struggles with their computing environment; the Laboratoryfor Advanced Systems Research of the University of Texas at Austin, specially toProf. Lorenzo Alvisi, for hosting me during my research visit in 2007; Dr. VolkerStrumpen of IBM and Prof. Keshav Pingali of the University of Texas at Austinfor their suggestions for the improvement of this work; and the VARIDIS researchgroup of the Politechnical University of Catalonia for providing access to their Feketeapplication.

I gratefully thank the following institutions for funding this work: the Depart-ment of Electronics and Systems of A Coruña for the human and material support;the University of A Coruña for nancing my attendance to some conferences; the

xviii

Xunta de Galicia (project PGIDIT04TIC105004PR); the Ministry of Science andInnovation of Spain (FPU grant AP-2004-2685 and projects TIN2004-07797-C02-02and TIN2007-67537-C03-02); and CYTED (Iberoamerican Programme of Scienceand Technology for Development; project 506PI0293).

Research is what I'm doing when Idon't know what I'm doing

Wernher Von Braun

Contents

Preface 1

1. Fault tolerance for parallel applications 7

1.1. Approaches for parallel fault tolerance . . . . . . . . . . . . . . . . . 8

1.1.1. Using intercommunicators . . . . . . . . . . . . . . . . . . . . 8

1.1.2. Modications to MPI semantics . . . . . . . . . . . . . . . . . 9

1.1.3. MPI extensions . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.1.4. Rollback-recovery . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2. Checkpoint-based approaches . . . . . . . . . . . . . . . . . . . . . . 11

1.2.1. Uncoordinated checkpointing . . . . . . . . . . . . . . . . . . 11

1.2.2. Coordinated checkpointing . . . . . . . . . . . . . . . . . . . . 13

1.2.3. Communication-induced checkpointing . . . . . . . . . . . . . 15

1.3. Log-based approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.3.1. Pessimistic log-based approaches . . . . . . . . . . . . . . . . 16

1.3.2. Optimistic log-based approaches . . . . . . . . . . . . . . . . . 17

1.3.3. Causal log-based approaches . . . . . . . . . . . . . . . . . . . 18

1.4. Implementation properties of rollback-recovery protocols . . . . . . . 18

xxi

xxii CONTENTS

1.4.1. Granularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.4.2. Transparency . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.4.3. Portability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.5. Existing checkpointing tools . . . . . . . . . . . . . . . . . . . . . . . 21

1.5.1. CoCheck . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.5.2. CLIP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.5.3. Porch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1.5.4. Egida . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1.5.5. Starsh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.5.6. MPICH-V2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.5.7. MPICH-GF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.5.8. PC3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.6. Proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.6.1. Spatial coordination . . . . . . . . . . . . . . . . . . . . . . . 25

1.6.2. Granularity and portability . . . . . . . . . . . . . . . . . . . 27

1.7. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2. CPPC Library 31

2.1. View . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.1.1. CPPC initialization and shutdown . . . . . . . . . . . . . . . 32

2.1.2. Variable registration . . . . . . . . . . . . . . . . . . . . . . . 34

2.1.3. Non-portable calls . . . . . . . . . . . . . . . . . . . . . . . . 35

2.1.4. Context management . . . . . . . . . . . . . . . . . . . . . . . 37

2.1.5. Open les . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

CONTENTS xxiii

2.1.6. Checkpoint le dumping . . . . . . . . . . . . . . . . . . . . . 39

2.1.7. Application restart . . . . . . . . . . . . . . . . . . . . . . . . 39

2.2. Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.3. Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.3.1. Façade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.3.2. Checkpointing layer . . . . . . . . . . . . . . . . . . . . . . . . 53

2.3.3. Writing layer . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2.3.4. Portability layer . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2.3.5. Utility layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.4. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3. CPPC Compiler 63

3.1. Compiler overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.2. Analyses and transformations . . . . . . . . . . . . . . . . . . . . . . 67

3.2.1. CPPC initialization and nalization routines . . . . . . . . . . 67

3.2.2. Procedure calls with non-portable outcome . . . . . . . . . . . 68

3.2.3. Open les . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3.2.4. Conversion to CPPC statements . . . . . . . . . . . . . . . . . 69

3.2.5. Data ow analysis . . . . . . . . . . . . . . . . . . . . . . . . 70

3.2.6. Communication analysis . . . . . . . . . . . . . . . . . . . . . 73

3.2.7. Checkpoint insertion . . . . . . . . . . . . . . . . . . . . . . . 78

3.2.8. Language-specic transformations . . . . . . . . . . . . . . . . 81

3.2.9. Code generation . . . . . . . . . . . . . . . . . . . . . . . . . . 82

3.3. Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

xxiv CONTENTS

3.4. Implementation details . . . . . . . . . . . . . . . . . . . . . . . . . . 86

3.4.1. Fortran 77 support . . . . . . . . . . . . . . . . . . . . . . . . 86

3.4.2. Sharing code between the C and Fortran 77 compilers . . . . . 88

3.4.3. AST analyzers . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3.5. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

3.6. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4. Experimental Results 95

4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

4.2. Compiler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

4.2.1. Analysis of the instrumented codes . . . . . . . . . . . . . . . 98

4.2.2. Compilation times . . . . . . . . . . . . . . . . . . . . . . . . 118

4.3. Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

4.3.1. State le sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

4.3.2. State le creation time . . . . . . . . . . . . . . . . . . . . . . 124

4.3.3. Checkpoint overhead . . . . . . . . . . . . . . . . . . . . . . . 131

4.3.4. Restart times . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

4.4. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

Conclusions and future work 141

Bibliography 145

List of Tables

2.1. Interface summary of the CPPC library . . . . . . . . . . . . . . . . . 33

2.2. CPPC conguration parameters . . . . . . . . . . . . . . . . . . . . . 45

3.1. Semantic roles used by the CPPC compiler . . . . . . . . . . . . . . . 66

4.1. Summary of test applications . . . . . . . . . . . . . . . . . . . . . . 96

4.1. Summary of test applications (continued) . . . . . . . . . . . . . . . . 97

4.2. Detail of loops selected by the shape-based threshold for NAS BT . . 100

4.3. Detail of loops selected by the shape-based threshold for NAS CG . . 102

4.4. Detail of loops selected by the shape-based threshold for NAS EP . . 102

4.5. Detail of loops selected by the shape-based threshold for NAS FT . . 104

4.6. Detail of loops selected by the shape-based threshold for NAS IS . . . 107

4.7. Detail of loops selected by the shape-based threshold for NAS LU . . 107

4.8. Detail of loops selected by the shape-based threshold for NAS MG . . 109

4.9. Detail of loops selected by the shape-based threshold for NAS SP . . 109

4.10. Detail of loops selected by the shape-based threshold for CESGA

CalcuNetw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

4.11. Detail of loops selected by the shape-based threshold for CESGA Fekete115

xxv

xxvi LIST OF TABLES

4.12. Detail of loops selected by the shape-based threshold for DBEM . . . 118

4.13. Detail of loops selected by the shape-based threshold for STEM-II . . 118

4.14. Statistics and compilation times for test applications . . . . . . . . . 120

4.15. Breakdown of compilation times for test applications . . . . . . . . . 121

4.16. Number of nodes and cores used for runtime tests for each application 122

4.17. Performance of the automatic variable registration algorithm . . . . . 124

4.18. Checkpoint le creation times (seconds) . . . . . . . . . . . . . . . . . 132

4.19. Runtime overhead caused by checkpointing . . . . . . . . . . . . . . . 135

4.20. Runtime overhead on large-scale applications . . . . . . . . . . . . . . 136

4.21. Restart times (seconds) . . . . . . . . . . . . . . . . . . . . . . . . . . 137

List of Figures

1. Design of the CPPC framework . . . . . . . . . . . . . . . . . . . . . 4

1.1. Domino eect in uncoordinated checkpointing . . . . . . . . . . . . . 13

1.2. Spatial coordination for non-blocking coordinated checkpointing . . . 26

2.1. Global design of the CPPC framework . . . . . . . . . . . . . . . . . 43

2.2. Pseudocode of the algorithm for nding the recovery line . . . . . . . 47

2.3. Data hierarchy format used for writing plugins . . . . . . . . . . . . . 52

3.1. Semantic information for the fopen and open functions . . . . . . . . 65

3.2. Example of a non-portable procedure call transformation . . . . . . . 68

3.3. Pseudocode for a le opening transformation . . . . . . . . . . . . . . 69

3.4. Pseudocode for a le closing transformation . . . . . . . . . . . . . . 69

3.5. First step: shape-based threshold . . . . . . . . . . . . . . . . . . . . 79

3.6. Second step: cluster-based threshold . . . . . . . . . . . . . . . . . . 80

3.7. Modications to a checkpointed Fortran 77 loop . . . . . . . . . . . . 81

3.8. Case study: original DBEM code . . . . . . . . . . . . . . . . . . . . 83

3.9. Case study: CPPC-instrumented DBEM code . . . . . . . . . . . . . 84

3.10. Case study: CPPC-instrumented DBEM code (cont.) . . . . . . . . . 85

xxvii

xxviii LIST OF FIGURES

3.11. AST analyzer implementation by instantiation of the method dispatcher 90

4.1. Checkpoint insertion for NAS BT . . . . . . . . . . . . . . . . . . . . 99

4.2. Checkpoint insertion for NAS CG . . . . . . . . . . . . . . . . . . . . 101

4.3. Checkpoint insertion for NAS EP . . . . . . . . . . . . . . . . . . . . 103

4.4. Checkpoint insertion for NAS FT . . . . . . . . . . . . . . . . . . . . 105

4.5. Checkpoint insertion for NAS IS . . . . . . . . . . . . . . . . . . . . . 106

4.6. Checkpoint insertion for NAS LU . . . . . . . . . . . . . . . . . . . . 108

4.7. Checkpoint insertion for NAS MG . . . . . . . . . . . . . . . . . . . . 110

4.8. Checkpoint insertion for NAS SP . . . . . . . . . . . . . . . . . . . . 111

4.9. Checkpoint insertion for CESGA CalcuNetw . . . . . . . . . . . . . . 113

4.10. Checkpoint insertion for CESGA Fekete . . . . . . . . . . . . . . . . 114

4.11. Checkpoint insertion for DBEM . . . . . . . . . . . . . . . . . . . . . 116

4.12. Checkpoint insertion for STEM-II . . . . . . . . . . . . . . . . . . . . 117

4.13. File sizes for NAS BT . . . . . . . . . . . . . . . . . . . . . . . . . . 125

4.14. File sizes for NAS CG . . . . . . . . . . . . . . . . . . . . . . . . . . 125

4.15. File sizes for NAS EP . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

4.16. File sizes for NAS FT . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

4.17. File sizes for NAS IS . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

4.18. File sizes for NAS LU . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

4.19. File sizes for NAS MG . . . . . . . . . . . . . . . . . . . . . . . . . . 128

4.20. File sizes for NAS SP . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

4.21. File sizes for CESGA Fekete . . . . . . . . . . . . . . . . . . . . . . . 129

4.22. File sizes for DBEM . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

LIST OF FIGURES xxix

4.23. File sizes for STEM-II . . . . . . . . . . . . . . . . . . . . . . . . . . 130

4.24. Summary of le sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

4.25. Maximum mean dumping times for test applications . . . . . . . . . . 133

4.26. Restart times for test applications . . . . . . . . . . . . . . . . . . . . 138

Preface

Parallel computing evolution towards cluster and Grid infrastructures has creatednew fault tolerance needs. As parallel machines increase their number of processors,so does the failure rate of the global system. This is not a problem as long asthe mean time to complete an application's execution remains well under the meantime to failure (MTTF) of the underlying hardware, but that is not always truefor applications with large execution times. Under these circumstances, users andprogrammers need a way to ensure that not all computation done is lost on machinefailures.

Checkpointing has become a widely used technique to obtain fault tolerance.It periodically saves the computation state to stable storage so that the applica-tion execution can be resumed by restoring such state. A number of solutions andtechniques have been proposed [22], each having its own pros and cons.

Current trends towards new computing infrastructures, such as large heteroge-neous clusters and Grid systems, present new constraints for checkpointing tech-niques. Heterogeneity makes it impossible to apply traditional state saving tech-niques which use non-portable strategies for recovering structures such as applicationstack, heap, or communication state.

Therefore, modern checkpointing techniques need to provide strategies for por-table state recovery, where the computation can be resumed on a wide range ofmachines, from binary incompatible architectures to incompatible versions of soft-ware facilities, such as dierent implementations for communication interfaces.

This work presents a checkpointing framework focused on the automatic insertionof fault tolerance into long-running message-passing applications. It is designed toallow for execution restart on dierent architectures and/or operating systems, alsosupporting checkpointing over heterogeneous systems, such as the Grid. It usesportable code and protocols, and generates portable checkpoint les while avoiding

1

2 PREFACE

traditional solutions (such as process coordination or message logging) which addan unscalable overhead.

Checkpointing evolution

Checkpointing techniques appeared as operating system (OS) services, usuallyfocused on recovery of sequential applications. Examples are KeyKOS [36], whichperformed system-wide checkpointing, storing the entire OS state, and the Sen-try [56], a UNIX-based implementation which performed checkpointing and journal-ing for logging of non-deterministic events on single processes. Sprite [45] dealt withthe process migration aspect of checkpoint-and-restart recovery for shared memorycomputers. Its focus was balancing the workload of a network of workstations bymigrating all the processes executing a single shared-memory application to idlemachines. Being implemented into the OS, these checkpointing solutions were com-pletely ad-hoc, with a lack of emphasis on portability. Their approach to operationeciency was based on achieving good I/O performance for storing the whole com-putation state, instead of reducing the amount of data to be stored.

The rst obvious disadvantage of OS-based implementations is the hard depen-dency between fault tolerance and the operating system of choice. Checkpointingfacilities, which were a common feature in earlier operating systems, gradually dis-appeared in the early 90s and were unavailable for popular environments such asUNIX, SunOS or AIX. The desire for exible solutions which could operate in dier-ent environments motivated the emergence of application level solutions (as opposedto system level solutions). In these tools, fault tolerance is achieved by compilingthe application program together with the checkpointing code, usually found in alibrary. Checkpointing solutions in this period were still transparent, storing theentire application state. Not being implemented inside the kernel they had to solveimportant problems when manipulating OS-dependent state. Examples are restor-ing process identiers or tracing open les. Also, they had to gure out ways torecover the application stack or heap. These issues made their codes still very de-pendent on specic operating system features. Usually, this forced developers torestrict the type of OS facilities used by the checkpointed programs. Examples ofapplication level, transparent tools include Libckpt [49], CATCH GCC [38] (a mod-ied version of the GNU C compiler), and the uniprocessor checkpointing facilityincluded with the Condor batch system [41].

PREFACE 3

Also in the mid-90s some non-transparent solutions tried to apply checkpoint-ing to distributed platforms. Their fundamental drawback was the lack of commonground regarding the interface for interprocess communication, which made these so-lutions tied to a specic and non-standard interface. Examples of these frameworksare Calypso [8] and extensions to Dome (Distributed Object Migration Environ-ment) [10]. In both cases the programming language used was an extension of C++with non-standard parallel constructs.

The adoption of MPI as the de-facto standard for parallel programming spawnedthe appearance of many MPI-based checkpointing tools in the last years. At rst,these used the transparent application level approach, sharing the same drawbacks astheir uniprocessor counterparts: lack of data portability and restriction of supportedenvironments, which here refers to the underlying MPI implementation. In fact,checkpointers in this category are generally implemented by modifying a previouslyexisting MPI library. Examples of these types of checkpointers are MPICH-GF [71]and MPICH-V2 [12], both implemented as MPICH drivers, thus forcing all machinesto run this specic implementation.

More recently, the advent of the Grid requires that checkpointers evolve towardsapplication level approaches that enable both data portability, by storing data usingportable representation formats, and communication-layer independence, by imple-menting the solution in a higher level of abstraction.

The CPPC Framework

As stated in the previous section, modern computing trends require portabletools for message-passing applications, focusing on providing the following funda-mental features:

1. OS-independence: checkpointing strategies must be compatible with any givenoperating system. This means having at least a basic modular structure toallow for substitution of certain critical sections of code (e.g. lesystem access)depending on the underlying OS.

2. Support for parallel applications with communication protocol independence:the checkpointing framework should not make any assumption as to the com-munication interface or implementation being used. Computational Grids in-clude machines belonging to independent entities which cannot be forced to

4 PREFACE

Figure 1: Design of the CPPC framework

provide a certain version of the MPI interface. Even recognizing the role ofMPI as the message-passing de-facto standard, the checkpointing techniquecannot be theoretically tied to MPI in order to provide a truly portable,reusable approach.

3. Reduced checkpoint le sizes: the tool should optimize the amount of databeing saved, avoiding dumping state which will not be necessary upon appli-cation restart. This improves performance, which depends heavily on state lesizes. It also enhances performance of the process migration in computationalGrids.

4. Portable data recovery: the state of an application can be seen as a structurecontaining dierent types of data. The checkpointing tool must be able torecover all these data in a portable way. This includes recovery of opaquestate, such as MPI communicators, as well as of OS-dependent state, such asthe le table or the execution stack.

The CPPC framework (ComPiler for Portable Checkpointing) provides all thesefeatures which are key issues for fault-tolerance support on heterogeneous systems. Itappears to the user as a runtime library containing checkpoint-support routines [54],together with a compiler which automates the use of the library. The global pro-cess is depicted in Figure 1. Upon runtime, the fault tolerant parallel applicationperforms calls to the routines provided by the CPPC Library.

Organization of this Thesis

The structure of this work is as follows. Chapter 1 describes the various alter-natives for fault tolerance in parallel applications. It surveys the existing rollback-recovery tools focusing on MPI applications. It also describes the CPPC proposal,outlining its most important characteristics and design decisions.

Chapter 2 provides an in-depth description of the design and implementation ofthe CPPC library following a top-down approach, starting at the application code

PREFACE 5

and descending through the various layers in which the library is divided. It alsofocuses on the modications that need to be performed to the original code in orderto achieve integration with CPPC.

In Chapter 3 the design and implementation of the CPPC source-to-source com-piler are covered. It also provides details on the design decisions taken to enablethe reusability of the platform. Finally, it includes a related work section coveringresearch focused on the static analysis and checkpoint insertion of programmingcodes.

Chapter 4 describes the thorough experimental evaluation of the entire CPPCframework, including results to assess the performance of both the compiler andthe checkpointing library. The experimental tests were performed on the Finis Ter-rae supercomputer hosted by the Galician Supercomputing Center (CESGA). Thisprovides a more accurate performance estimation than using highly controlled andclosed environments available to a particular research group only.

Finally, the work is concluded by outlining its most relevant aspects and dis-cussing future research lines.

Chapter 1

Fault tolerance for parallel

applications

The need for an evergrowing computational power has always been one of the fun-damental driving forces of computer evolution, motivated by emerging requirementsin such application elds as sciences, engineering, or even leisure. The traditionalapproach for this evolution has been the enhancement of already-known technolo-gies: building Von Neumann machines with faster processors and larger amounts ofmemory. This approach is not always feasible, however. There are physical limitsimposed by several factors, such as the machines in charge of circuit integration,or the amount of energy the circuits are able to dissipate as heat. There are alsoeconomic constraints: the cost of new developments does not achieve return-on-investment for certain projects. These limitations are one of the most importantreasons for the popularization of parallel computing: an approach that seeks tobreak down a problem into subtasks which can be solved in a concurrent fashion.

Parallel applications are a special target for fault tolerance. Interprocess commu-nications generate dependencies between processes participating in a parallel execu-tion. Therefore, fault tolerance mechanisms cannot be applied individually to eachprocess in an independent way. Simply recovering a failed process can break theestablished dependencies, thereby creating inconsistencies at the application level.

Conceptually dierent approaches to solving the parallel fault tolerance problemhave been developed. Although rollback-recovery is the most widely used, a briefsurvey of other options is included in order to completely understand its advantagesover its competitors. At the end of this chapter, Section 1.6 gives an overview of

7

8 Chapter 1. Fault tolerance for parallel applications

the fault-tolerance tool introduced in this work.

1.1. Approaches for parallel fault tolerance

This chapter focuses on fault tolerance approaches for MPI applications [31]. Thereason is MPI's status as de-facto standard. However, there is no real limitation forapplying the covered techniques to other frameworks, like PVM (Parallel VirtualMachine) [65].

1.1.1. Using intercommunicators

For applications organized in a master-slave fashion, the failure of a client doesnot signicantly aect the server or the other clients that are able to continue execut-ing properly. This structure is robust because communications are always two-sided,and one of the sides is able to easily detect that the other one has failed. Therefore,it is a good context for introducing fault tolerance.

In MPI, the structure which models this kind of two-sided communications isthe intercommunicator. Processes in an intercommunicator are organized into twogroups, and all communications take place between processes in one group and pro-cesses in the other. In master-slave applications, the master manages a set of taskswhich are submitted to slaves that carry them out. These slaves return the resultsto the master when the job is done, and ask for a new task to perform. All com-munications take place between a slave and the master, and there are no collectivecommunications nor communications between two dierent slave processes.

In order for the master to detect a slave's failure, it can simply replace the defaultfailure handler MPI_ERRORS_ARE_FATAL with a new one, which reassigns the task thefaulty slave was performing to a dierent one. Tasks are usually small, and thereforethe slave state at the time of failure can be forgotten, simply starting the task over.

In case the master fails, it is necessary to use any of the other approaches in thissection for recovering its state. The use of checkpoint-based rollback recovery (seeSection 1.1.4) is usually very ecient, since the master's state consists basically oftask specications, their current statuses, and results of completed ones.

1.1 Approaches for parallel fault tolerance 9

1.1.2. Modications to MPI semantics

MPI objects and function semantics can be modied so that its behavior ismade inherently fault tolerant. For instance, collective communications could behavedierently in the presence of failed processes, or certain pieces of data could beredistributed upon failure detection.

One such approach is FT-MPI [24], a fault tolerant MPI implementation thatcan survive failures, while oering the application developer a range of recoveryoptions. The approach in this case is to extend MPI semantics, substituting the twoMPI communicator states (valid or invalid) by an array which can be simplied toOK, PROBLEM, FAILED. The introduction of an intermediate state allows FT-MPIto deal with communications in the presence of errors in a fault-tolerant way.

When a non-recoverable error in a process is detected, its communicator needsto be modied, either making it smaller (redistributing ranks), leaving a blank tobe lled later, or creating a new process to ll in the blank left by the failed one.

While interesting from a scientic point of view, this approach sacrices the veryreason for MPI existence: the standard. Fault tolerance remains a property of theMPI implementation, and not of the application code, which will behave dierentlywhen executed on standard MPI implementations and the semantically modiedone.

1.1.3. MPI extensions

Instead of modifying function and object semantics, new ones are introduced toprovide fault tolerance. This approach presents the same problem as did semanticmodications, only worse. When extending the MPI interface, the application codewill not even compile if using a non-extended MPI implementation. Besides, extend-ing MPI is no better than developing isolated, orthogonal fault tolerance libraries.This is far more interesting, since it may provide fault tolerance capabilities withoutrestricting the communication environment. One example of extensions to MPI isMPI/FT [9].

10 Chapter 1. Fault tolerance for parallel applications

1.1.4. Rollback-recovery

The most widespread approach to fault tolerance is rollback-recovery. It is basedon periodically saving the state of the execution to stable storage, recovering it incase of failure. A system is considered to have been correctly recovered if the ex-ternally visible behavior of the distributed system is equivalent to some failure-freeexecution [62]. Note that it is not required that the recovered state has occurred ina previous execution. Rather, it is sucient that it might have happened. There aremany dierent approaches to rollback-recovery, each of them very dierent in theway it implements this basic idea. Although many tools tie themselves to specicsystems at the implementation level, a fundamental advantage of rollback recoveryis that it is essentially independent of the programming paradigm or the commu-nication system being used. Of course, applying it to message-passing applicationsimplies performing certain tasks to ensure the consistency of the global system,which should not need to be done when dealing with sequential applications. Aconsistent system state for a message-passing application is described by Chandyand Lamport as one in which every message reected as received by a process'sstate is reected as sent in the corresponding sender process's state [18]. This def-inition allows for a message to be reected as sent by a process, but not receivedby its recipient. This is not considered an inconsistent state, since it might happenin a regular execution. However, the message is considered in-transit [22]. If theintended receiver of an in-transit message has failed and is recovered on a systembuilt on reliable communication channels, the rollback-recovery protocol needs toensure that the message is recreated and delivered to its recipient. If the system isbuilt on non-reliable channels, the failure of the receiver is indistinguishable fromthe failure of the channel itself, and therefore the delivery of the message will behandled by the communications system.

There are two core approaches to rollback-recovery: checkpoint-based and log-based. The former ensures that the processes' state is recovered correctly and con-sistently up to the last recovery line, which is dened as the most recent consistentset of checkpoints [52]. However, they do not guarantee that the execution of partsof the code that had already been executed prior to the failure, but after creating thecheckpoint used for recovery, will be deterministically re-executed exactly as theyoriginally were. Log-based approaches, on the contrary, create and log to stable stor-age information about non deterministic events, which allows them to recreate suchevents after recovery exactly as they played out in the pre-failure execution. Theseapproaches rely on the so called piecewise deterministic (PWD) assumption [62]: all

1.2 Checkpoint-based approaches 11

nondeterministic events that a process executes can be identied, and the informa-tion necessary to replay each event during recovery can be logged and encoded intuples called the event's determinant. Although log-based approaches are usuallycalled message-logging due to the rst systems which proposed and implementedthese techniques, the log is not restricted to interprocess communications. Moretypes of events, such as generation of random variables, might be logged dependingon the actual range of nondeterministic events covered under the PWD for eachdierent system. This is an implementation issue. Note that sending a message isnot a nondeterministic event.

The next section is devoted to checkpoint-based rollback recovery techniques.Section 1.3 covers log-based approaches.

1.2. Checkpoint-based approaches

As explained in the previous section, approaches based on checkpointing are ableto correctly and consistently recover an application's state, but they do not rely onthe PWD assumption. Therefore, pre-failure execution cannot be deterministicallyregenerated after a rollback. These kinds of approaches are, therefore, not a goodchoice when dealing with applications that should preserve their pre-failure behaviorunchanged, such as applications with frequent interactions with the outside world.

Checkpoint-based approaches are classied into three categories: uncoordinated,coordinated and communication-induced.

1.2.1. Uncoordinated checkpointing

In uncoordinated checkpointing, each process is allowed to take its local check-points independently from others. This has advantages. First, there is no need tocomply with any coordination protocol that introduces execution overhead. Second,each process can take its local checkpoints when it better ts its needs. For instance,it can do so when the process predicts it is going to be less expensive. However, nd-ing recovery lines in uncoordinated checkpointing is a dicult task. The approach issusceptible to domino eect [52], and since processes checkpoint independently thereis the possibility of creating useless checkpoints, which are those that are never partof a valid recovery line. This approach also needs a fairly complex garbage collection

12 Chapter 1. Fault tolerance for parallel applications

scheme. These problems are further described in the following subsections.

• Finding valid recovery lines

When a process fails using uncoordinated checkpointing, it must be rolled backto a previous checkpoint. However, doing so might break interprocess dependenciesintroduced by communications. If no further actions are taken, just simply selectingthe most recent checkpoint in the pool will not be enough, since the recovered statemay be inconsistent. In order to nd out whether a state is inconsistent, the failedprocess needs information about the global dependencies in the application.

The way uncoordinated approaches collect these dependencies is by piggybackingits status into messages sent to other processes [11]. Processes receiving a messageare able to detect that their current state depends on the sender's process statewhen it sent the message. When a process fails, it asks all other processes to sendit their saved dependency information. Using this information, it is able to builda graph for nding a valid recovery line. This involves rolling back other processesas well to recover a consistent state. Two dierent, equivalent methods for buildingand analyzing such a graph were proposed [11,70].

• Garbage collection

Garbage collection is a very important process in uncoordinated checkpointing.Since processes create checkpoints independently, eventually some checkpoint leswill be never used again. Precisely, for each process p, all les older than theone in the most recent valid recovery line are obsolete, and can be purged. Thesame algorithm used for nding valid recovery lines is used for determining whichcheckpoints can be safely purged. This is an expensive process.

• Domino eect

One serious consequence of uncoordinated checkpointing and the algorithms fornding valid recovery lines is that their very existence is not guaranteed. Sincecheckpoints are taken in an independent way, situations may occur in which no validrecovery lines can be found, because interprocess dependencies force all processesto roll back to the very beginning of the application execution. One such situationis depicted in Figure 1.1. Each mi denotes a message being sent, while each cj,k

1.2 Checkpoint-based approaches 13

denotes the k-th checkpoint in process j. When p3 fails, it rst considers rollingback to checkpoint c3,2. Since message m8 imposes a dependency on p2, it would beforced to roll back to c2,2. This, due to m7, would force p1 to roll back to c1,2. Theanalysis moves forward, and the protocol in charge of nding a valid recovery linediscovers that the only valid solution is to roll back all processes to their respectivebeginnings.

Figure 1.1: Domino eect in uncoordinated checkpointing

1.2.2. Coordinated checkpointing

The idea behind coordinated checkpointing is to ensure that all processes taketheir checkpoints in a way that ensures that no inconsistent state recovery can takeplace. There are several ways in which this can be done. The most obvious oneis to stop all application progress while the checkpoint is taken [66]. An initia-tor, which may be an application process or an external entity, sends a broadcastmessage requesting that processes checkpoint. Upon receiving this message, eachprocess ushes its local communication channels and creates its local checkpointbefore resuming execution. This process may lead to signicant overheads due toits blocking nature.

In order to reduce the eects of coordination, non-blocking coordinated schemeswere proposed. The most studied protocol to date has been the distributed snap-

shots protocol by Chandy and Lamport [18]. The protocol starts with the initiatorbroadcasting a checkpoint request. Upon receiving the request, each process takes

14 Chapter 1. Fault tolerance for parallel applications

its checkpoint and rebroadcasts the request before sending any more applicationmessages. The goal is to ensure that processes do not receive any message thatcould make the global state inconsistent. However, in order for this to be true,the communication channels need to be FIFO and reliable. When working withnon-FIFO channels, each process that has just taken its local checkpoint could pig-gyback the checkpoint request on all further sent application messages, to counterthe possibility of the request being received out of order. Another option is to makeeach process piggyback its current epoch1 in all sent messages. This automaticallynoties other processes that a checkpoint has been taken upon receival of a messagewith an epoch number higher than their own.

If the channel is reliable, coordinated checkpoint techniques ensure that no in-consistencies take place upon recovery. However, in-transit messages still have to betaken into account. If the communications channel is reliable, processes need to logmessages received after taking its local checkpoint but sent before the sender tookits own. If the communication channel were non-reliable, then in-transit messagesare equivalent to messages lost in channel errors, and therefore no further actionsneed to be taken.

Using coordinated checkpointing, not only the failed processes must roll back, butall processes involved in the computation have to do so as well. This ensures that allevents which aected the failed process after the checkpoint was taken are replayed.In this case, nding recovery lines is a very easy operation. Since all processes taketheir checkpoints in a coordinated way, the last valid recovery line is composed ofthe latest common checkpoints. Due to this, the garbage collection algorithm ispretty straightforward, and simply consists of removing the oldest checkpoints whennewer ones are created.

Another variant involves taking checkpoints based on a coordinated clock [21].Other authors proposed notifying only those processes which transitively depend onthe initiator [35]. The result is that not all processes participate in all checkpoints,but rather, only those that present dependencies among themselves.

1In checkpointing terminology, an epoch is the interval between two given checkpoint actions.Therefore, the computation can be divided in a series of numbered epochs.

1.3 Log-based approaches 15

1.2.3. Communication-induced checkpointing

Communication-induced checkpointing is based on piggybacking dependenciesinformation on application messages so that the domino eect and the creationof useless checkpoints can be avoided. Processes are not coordinated, but ratheranalyze the state of the global computation they have knowledge of and make de-cisions about its local checkpoints in two dierent ways. First, they may take localcheckpoints based on their own execution state as in uncoordinated checkpoint ap-proaches. Second, they also take forced checkpoints depending on the global depen-dencies state to avoid the domino eect and the existence of useless checkpoints.

The idea on which communication-induced is based are Z-paths and Z-cycles [44].It has been proven that a checkpoint is useless if and only if it is part of a Z-cycle.These approaches piggyback protocol information in application messages so thatprocesses are able to detect when they should take a forced checkpoint to avoid theexistence of useless ones. There are two basic approaches: model-based, which detectZ-cycles and force checkpoints to eliminate them; and index-based, which timestampeach local checkpoint on each process and piggyback these timestamps on applicationmessages so that each process decides whether or not to take forced checkpoints,analyzing the timestamps received by other processes. These two approaches havebeen proven to be fundamentally equivalent [33].

Communication-induced checkpointing is theoretically more ecient and scalablethan coordinated checkpointing. However, this technique may indeed create check-points that will not be used, and is extremely dependent on the message-passingpattern of the application. This leads to unpredictable checkpoint rates, making theapproach dicult to use in practice [23].

1.3. Log-based approaches

Log-based approaches consider the execution of a parallel application as a se-quence of deterministic state intervals. Each interval begins with the execution ofa nondeterministic event, modeled as the reception of a nondeterministic message.Nondeterministic events include application messages and other forms of nondeter-minism such as generation of random numbers or reception of user events. By loggingnondeterministic events and relying on the PWD assumption, log-based techniquesguarantee that the recovered execution will exactly recreate the nondeterministic

16 Chapter 1. Fault tolerance for parallel applications

events that had already taken place. Since deterministic events generated by thefailed process (e.g. sent messages) will be recreated exactly as they had been in thepre-failed execution, non-failed processes are guaranteed not to become orphaned bythe roll back of the failed processes and are never required to roll back themselves, asopposed to checkpoint-based rollback-recovery. This is formalized in the no-orphansconsistency condition [7]:

Let Depend(e) be the set of processes that are aected by a nondeterministicevent e.

Let Log(e) be the set of processes that have logged the determinant of evente in their volatile memory.

Let Stable(e) be a predicate that is true if e's determinant has been logged tostable storage.

∀e : ¬Stable(e)→ Depend(e) ⊆ Log(e)

This means that, for any event e, if it is not logged in stable storage, then allprocesses depending on it must have it available in their volatile memory. Indeed,if a process depending on an event e could not access it via stable storage norvolatile log, this process would be an orphan process, since it would depend on anon-recoverable nondeterministic event.

These kind of approaches periodically create uncoordinated checkpoints in orderto reduce the amount of re-execution needed upon recovery.

There are basically three types of log-based protocols for guaranteeing that theno-orphans consistency condition is true: pessimistic, optimistic and causal. Theyare described in the following subsections.

1.3.1. Pessimistic log-based approaches

Pessimistic approaches synchronously log every nondeterministic event as soon asit is generated, not allowing execution to resume until it is available in stable storage.Therefore, they implement a stronger condition than the no-orphans consistency one.It may be logically expressed as:

1.3 Log-based approaches 17

∀e : ¬Stable(e)→ |Depend(e)| = 0

That is, if a nondeterministic event has not been logged to stable storage, thismeans no process depends on it.

These approaches have a main disadvantage: the overhead of synchronously log-ging all received nondeterministic events. In order to reduce this overhead, whichmakes the approach highly non-scalable, techniques such as using special logginghardware [2] are used. Another option is delaying the logging until the processcommunicates itself with any other process, an approach known as sender-basedpessimistic logging [34]. Depending on the communication patterns of the applica-tion, this may log more than one event at the same time, therefore reducing I/Ooverhead. This may be a problem when using reliable channels. If the process failsbefore logging the received messages to stable storage, these will not be available toit upon recovery.

In terms of garbage collection, these protocols require only the last stored check-point to be kept for each process.

1.3.2. Optimistic log-based approaches

These protocols asynchronously log nondeterministic events [62]. Each processkeeps its determinants in a volatile log, which is committed to stable storage peri-odically. These kinds of protocols have a good failure-free performance, althoughthey do not validate the no-orphans condition. Indeed, if a process were to fail,having determinants in its volatile log not yet committed, all processes dependingon events generated by the failed process would become orphans, which requiresthem to roll back. Thus, optimistic log-based approaches are very similar to unco-ordinated checkpointing in case that a process fails without having committed anydeterminant to stable storage. This results in complex recovery protocols, as wellas in a complex garbage collection. Moreover, if eventual stable log of determinantsis not guaranteed, the approach is subject to the domino eect.

18 Chapter 1. Fault tolerance for parallel applications

1.3.3. Causal log-based approaches

These protocols combine the advantages of both pessimistic and optimistic ap-proaches. They are based on piggybacking the contents of each process's volatile login messages it sends, so the information is available on other volatile logs. Therefore,surviving processes are able to help failed processes replay their pre-failed executionin a deterministic way. By doing so, it is guaranteed that no orphan processes mayexist, and therefore no processes are forced to roll back other than failed ones. Theirdrawback is the communications overhead introduced by the piggybacking of localdeterminants on sent messages.

1.4. Implementation properties of rollback-recovery

protocols

Whether the protocol used is checkpoint-based or log based, all rollback-recoveryprotocols create checkpoint les as their fundamental recovery unit. It is the onlydata available to processes in checkpoint-based approaches, and it allows log-basedprotocols to eciently recover after a failure.

There are important properties associated with the actual way of creating stateles that aect the performance and capabilities of real rollback-recovery tools. Wehighlight three such properties: granularity, transparency and portability.

1.4.1. Granularity

This property determines the amount of data that the fault tolerance approachstores as part of the checkpoint les. In particular, it determines how much data,compared to the total application data, the technique stores. There are two fun-damental approaches to this; full checkpointing, consisting in storing the entireapplication state, including structures such as the application stack or heap; andvariable-level checkpointing, which identies and selects variables which are neededfor restart and stores only those in the checkpoint les. These are sometimes re-lated to the level of abstraction at which the rollback-recovery is implemented. Ifit is implemented on a system level, that is, on a higher abstraction level thanthe application itself, it does not have any knowledge of the application internals.

1.4 Implementation properties of rollback-recovery protocols 19

Therefore, a variable-level analysis is not possible, and it must store the applicationstate entirely. Other approaches, implemented at application level, have access tothe application code and can modify it in order to introduce fault tolerance. Thesehave access to internal application information and can use it to their advantage.Thus, all system-level rollback-recovery tools must perform full checkpointing, whileapplication-level checkpointers can choose whether to opt for variable-level check-pointing instead.

The fundamental advantage of full checkpointing is that it treats an applicationas a black box, without knowledge of its internals or previous analysis. It is, there-fore, completely transparent for the programmer (see Section 1.4.2), as it can beimplemented at the operating system level or even in the hardware. However, ithas two important drawbacks. First, the literature has shown that the actual statewriting to stable storage is the largest contribution to the overhead of checkpoint-ing [47]. Storing all application data will have a higher associated cost than storingjust necessary data. Not only that, but transferring bigger checkpoint les over anetwork will also have a higher cost. Second, checkpoints will lack portability (seeSection 1.4.3). Variable-level checkpointing, on the other hand, is able to obtainbetter performance and may create portable checkpoint les depending on its im-plementation. The drawback is the need for complex analyses of the applicationcode in order to identify the state that needs to be stored.

• Incremental checkpointing

An optional improvement in both full and variable-level checkpointing is the useof incremental checkpointing [4,25,29,50]. Under this approach, there are two typesof checkpoints: full and incremental. Full checkpoints contain all data that is to bestored, while incremental checkpoints store only the data that has changed since thelast checkpoint. The recovery process uses the most recent available full checkpoint,and then orderly applies the changes reected in the incremental ones to completelyreconstruct the process state. Creating more incremental checkpoints between eachfull checkpoint improves the failure-free performance of the approach. Reducing itimproves recovery performance. A compromise must be reached, taking into accountthe failure rate of the system.

When incremental checkpointing is applied to full checkpointing, it is usuallydone at the page level, using hardware mechanisms to check which pages havechanged since the last checkpoint, and must be written in the incremental state

20 Chapter 1. Fault tolerance for parallel applications

le. When applied at the variable-level, some kind of algorithm to detect changesto individual variables must be applied. Therefore, incremental checkpointing willhave a higher impact in terms of runtime overhead when applied to variable-levelsystems.

1.4.2. Transparency

This property refers to how the user perceives the fault tolerance solution. Itis loosely related to granularity, since it might be clear that a full checkpointingscheme will be completely transparent for the users, while a variable-level one willrequire them to provide application-specic information. However, there are hybridapproaches that do not follow that rule. Some tools perform full checkpointing butstill require the user to select where checkpoints must be taken. Similarly, variable-level approaches may employ code analysis and transformations to automaticallyextract the information, behaving in a transparent way.

1.4.3. Portability

We say a checkpointing technique is portable if it allows the use of state les torecover the state of a failed process on a dierent machine, potentially binary incom-patible or using dierent operating systems or libraries. The basic condition that hasto be fullled in order to achieve potential portability is not to store any low-leveldata along with the process state. Therefore, all full checkpointing approaches arenon-portable. Note that not any variable-level approach will be portable, though.Some implementations sacrice portability for simplicity. For instance, one imple-mentation may store the program counter along with the application data in orderto simplify restart.

The second condition for a checkpointer to be portable is that all data is storedin a portable format, so conversions may be made in case they are necessary forrecovering the process state on a binary incompatible machine.

1.5 Existing checkpointing tools 21

1.5. Existing checkpointing tools

A number of tools for checkpointing have been proposed in the literature, eachone with its own approach. This section describes the most relevant ones, detail-ing what they miss, in our opinion, in order to be usable on modern computingframeworks.

1.5.1. CoCheck

CoCheck [61] is a full checkpointing library for parallel applications based inCondor [40]. It uses a coordinated checkpoint-based approach, with messages re-questing checkpoints acting as initiators. It works under the assumption of FIFOchannels. Upon reception of a checkpoint request, each process stores the state ofits communication buers and sends a request via all its communication channels.Therefore, any in-transit messages will be in the communication buers which arestored along with the process state, removing the possibility of orphan processes.No inconsistent messages exist since the approach is coordinated, and processes arenot allowed to send any new messages until their local checkpoint is completed.

The drawbacks of this approach are its coordinated nature, which hampers itsscalability, and the fact that it uses full checkpointing, which makes it unable towork on heterogeneous environments and also results in lower eciency. Besides,CoCheck depends on tuMPI, a specic MPI implementation. Therefore, the codewill not be portable to machines without tuMPI support.

1.5.2. CLIP

CLIP (Checkpointing Libraries for the Intel Paragon) [19] is focused on check-pointing parallel applications written for Intel Paragon architectures. This is aMIMD multicomputer that uses MPI or the Intel NX libraries for message-passingcommunications. CLIP implements ecient and simple solutions, but is heavily tiedto the Paragon architecture.

CLIP uses full checkpointing. However, it allows for dead memory regions to beexcluded from the checkpoint le, which improves eciency. It requires the user tospecify checkpoint locations in the application code. This is required because the

22 Chapter 1. Fault tolerance for parallel applications

authors did not wish to modify the Paragon kernel, and so they needed to ensurethat checkpointing does not take place inside a communication or I/O call.

It implements the same coordination technique seen in CoCheck: all processesare coordinated, FIFO channels are assumed and communication buers ushed.These are stored at userspace, and communication calls are substituted for CLIPversions which check that the requested message is not on the userspace buers, anddelegate their execution on the original version if it does not nd them. This solu-tion, although it is implemented on top of the MPI implementation, does require thelibrary to have knowledge about how the communication buers are implemented,and is therefore architecture-dependent and not applicable in a heterogeneous envi-ronment.

Portability is obviously not a design goal since the architecture is completelyParagon-centric.

1.5.3. Porch

Porch [51] is a source-to-source compiler that translates sequential C programsinto semantically equivalent C programs which are capable of saving and recoveringfrom portable checkpoints. The user inserts a call to a checkpoint routine andspecies the desired checkpointing frequency. A compiler then makes source-to-source transformations to instrument the operation. Its data hierarchy consists of ashadow stack. This is a copy of the original runtime stack, which is built by visitingeach stack frame and saving its local variables, identied at compile time. At thesame time, each active function is identied and the call sequence is stored. Thecomputation state is recovered by replaying the original call sequence and restoringeach stack frame using the shadow stack's contents. The shadow stack is stored ina portable format, called UCF (Universal Checkpoint Format).

Although this is a sequential approach, it is of fundamental importance in thecontext of this thesis, since it introduces fundamental ideas and techniques for check-pointing portability.

1.5.4. Egida

Egida [53] is a fault tolerance framework able to synthetize an implementationof the desired rollback-recovery protocol by interpreting a specication language to

1.5 Existing checkpointing tools 23

express its details. It facilitates rapid implementation of fault tolerance mechanismsfor MPI applications. Although it does not depend on any specic MPI implemen-tation, it implements the fault tolerance at the communication level. Therefore, lowlevel changes to the MPI layer are needed before an application can take advantan-tage of the synthesized modules.

1.5.5. Starsh

In Starsh [5], each MPI node runs a separate Starsh daemon, which commu-nicates with other daemons via the Ensemble group communication system [32].These daemons are responsible for spawning application processes, keeping track ofapplications health, managing the conguration and settings of the cluster, commu-nicating with clients, and providing the hooks necessary to provide fault tolerance.Its modular system enables the use of dierent checkpoint-based protocols, namelyuncoordinated and coordinated forms of checkpointing.

Since the checkpointing mechanisms is implemented at the OCaml virtual ma-chine, and given that OCaml is platform-independent, it is able to operate in hetero-geneous clusters despite doing full checkpoints. However, the use of OCaml is alsothe biggest drawback of the approach. Interpreting bytecode does is less ecientthan running native code. Doing full checkpoints also hinders performance. Besides,it is limited to host languages interpretable by the OCaml VM.

1.5.6. MPICH-V2

MPICH-V2 [12] is a communications driver for MPICH. It provides transparentfault tolerance by using the Condor checkpoint and restart capabilities. Therefore,it uses full checkpointing with no support for portability and heterogeneous envi-ronments.

This is a sender-based pessimistic logging approach. Each process stores a uniqueidentier and a timestamp for each message it receives. When it needs to rollback,it checks this registry in order to identify the messages that need to be replayed,and asks the original sender to do so. In addition to sending these messages, thesender process knows that the receiver will never ask for any previous message, andperforms garbage collection on its sending log.

Although the process is ecient, since it involves no process coordination, the

24 Chapter 1. Fault tolerance for parallel applications

scalability is hindered by the log-based approach. Performing full checkpointing doesnot help eciency nor portability. The fact that it is implemented as an MPICHdriver forces all machines to implement it in order to obtain fault tolerance.

1.5.7. MPICH-GF

MPICH-GF [71] is an extension to MPICH-G2 (the Globus communicationsdriver in MPICH) which adds fault tolerance. It is a transparent approach usingfull checkpointing. It uses coordinated checkpointing by using FIFO channels andushing communication buers upon receiving a checkpoint request from the initia-tor. It implements collective communications through point-to-point actions, whichis unecient. In fact, its overhead gures range between a 10% for the instrumen-tation and 20% if checkpointing is introduced.

1.5.8. PC3

The C3 system (Cornell Checkpoint Compiler) [1315] is the base for PC3 (Por-table C3) [26]. The user must insert checkpoint locations and a compiler is in chargeof orchestrating fault tolerance through full checkpointing. It is implemented on topof MPI, and therefore does not need any specic implementation to be used in orderto obtain fault tolerance.

In order to obtain consistency, it uses a non-blocking coordinated protocol, piggy-backing information into sent messages. Since it is on top of MPI, it cannot assumethe communications channel to be FIFO, since an application may decide to receiveits messages out-of-order. In order to solve this problem it implements a modiedversion of the Chandy-Lamport algorithm, which is also heavier. Particularly, itconverts collective communications into point-to-point ones, further damaging thescalability of the approach.

This approach is currently restricted to C applications.

1.6. Proposal

This work introduces a compiler-assisted checkpoint-based tool for portable check-pointing in parallel environments, called ComPiler for Portable Checkpointing (hence-

1.6 Proposal 25

forth, CPPC). It is composed of two cooperating systems: a library providing check-pointing routines for parallel applications, described in Chapter 2, and a compilerfor automating the code transformations necessary in order to take full advantage ofthe library, described in Chapter 3. This design combines some of the ideas exploredin this chapter with new ones focusing on achieving better qualitative and quanti-tative results. This section describes its most important properties, techniques andprotocols in terms of the properties described in this chapter.

1.6.1. Spatial coordination

As discussed in Section 1.2.1, uncoordinated checkpoint-based approaches presenta clear advantage: the lack of coordination overheads. However, they are also sub-ject to the domino eect, and are able to create useless checkpoints, that neverbecome part of valid recovery lines. It would be desirable to retain the eciencyand scalability obtained from not adopting complex runtime protocols while, at thesame time, guaranteeing the execution progress in the presence of failures. One wayto do so is to use spatially coordinated checkpointing. This is a form of coordinatedcheckpointing that, instead of taking all checkpoints at the same time forming aconsistent global state, focuses on SPMD codes and ensures that all checkpoints arecreated at the same code locations in a way that guarantees the consistency of therecovered state.

The basic dierence between parallel and sequential applications in terms ofconsistent recovery is the existence of dependencies imposed by interprocess com-munications. If a checkpoint is placed in the code between two matching communi-cation statements, an inconsistency would occur upon recovery, since the rst onewill not be executed. If it is a send statement, the message will not be resent andbecomes an in-transit message, also called in-ight, missing, or late. If it is a receivestatement, the message will not be received, becoming an inconsistent message, alsocalled ghost, orphan, or early.

The proposal of spatially coordinated checkpointing implies identifying, at com-pile time, code locations at which the non-existence of in-transit, nor inconsistentmessages is guaranteed. An example is shown in Figure 1.2. Let us dene a safe

point as a point in the code were it is guaranteed that no in-transit nor inconsistentmessages can exist. Let us assume that:

All checkpoints are placed at safe points.

26 Chapter 1. Fault tolerance for parallel applications

Figure 1.2: Spatial coordination for non-blocking coordinated checkpointing

All processes take the same number of checkpoints at the same safe points inthe code.

It is clear that, under these conditions, all checkpoints will take the same numberof checkpoints during their execution. Given two checkpoints ci,x, cj,y, let us denethe consistency relation, ci,x ∼ cj.y, as a binary relation that is true whenever thereare no in-transit, nor inconsistent messages between processes i and j between theirx-th checkpoint and y-th checkpoint respectively. Then, by the denition of safepoint:

∀i, j, x : ci,x ∼ cj,x (1.1)

Let us dene the set of all checkpoints taken by each process i, Ω(i). Finding avalid recovery line is as simple as nding the checkpoint index x that veries:

∀i : ci,x ∈ Ω(i) ∧ @y/ (∀j : cj,y ∈ Ω(j) ∧ y > x) (1.2)

That is, a checkpoint with index x exists in the set of checkpoints available forrecovering all processes, and it is the greatest index to fulll that condition. ByEquation 1.1, the set ci,x, i = 1, . . . , n will form a valid recovery line for allprocesses. According to Equation 1.2, it is also the most recent one. The garbagecollection algorithm will be equally simple. Processes can periodically communicate,in an asynchronous way, their most recently created checkpoint. By simply storing

1.6 Proposal 27

the minimum received value for all processes, a process i may determine that anycheckpoint with an index lower than the minimum will not be part of any validrecovery line in the future.

This approach has several advantages: it rules out both the possibility of thedomino eect, and the creation of useless checkpoints, without the need for anyspecic protocol at runtime; checkpoints are taken in a completely scalable way, sincethe number of processes does not aect in any way how chekpoints are created; noassumptions are made about the properties of the communications channel, whichcould be unreliable and/or non-FIFO without aecting the protocol.

The places at which checkpoints are to be created are statically determined bythe compiler, which needs to perform a static analysis to identify safe points and alsoan analysis to determine which of the detected safe points are adequate locationsfor checkpointing. The internals of these analyses are described in Chapter 3.

1.6.2. Granularity and portability

One of the most important design goals for CPPC is portability. With thisin mind, the possibility of performing full checkpointing is ruled out, since it isinherently non-portable. CPPC uses a variable level approach, storing only uservariables in state les. Furthermore, it only stores live variables, that is, those havingvalues which are necessary for state recovery. By restricting the amount of saveddata, checkpoint les are made smaller and so checkpointing overhead decreases.This also improves the performance of network transfers, if necessary.

The use of portable storage formats guarantees a consistent recovery of vari-able data in binary incompatible architectures. However, upon restart not only uservariables need to be recovered, but also non-portable state created in the original ex-ecution, such as MPI communicators, virtual topologies or derived data types. Thisintroduces the need for a restart protocol that regenerates the original non-portable,non-stored state. CPPC uses selective code re-execution to achieve complete appli-cation state recovery. Therefore, non-portable state is recovered by the same meansoriginally used to create it, making a CPPC application just as portable as theoriginal one: variables are saved using portable formats, while non-portable state isrecreated using the original code.

A piece of application code is dened as Required-Execution Code (REC) if itmust be re-executed at restart time to ensure correct state recreation. The recovery

28 Chapter 1. Fault tolerance for parallel applications

process consists of the selective and ordered re-execution of such blocks of code.There are six types of RECs, that correspond to:

Initialization of the CPPC library.

Portable state recovery: done by the CPPC registration routines.

Non-portable state recovery: performed through re-execution of procedureswith non-portable outcome.

Restart completion tests: carried out by the checkpointing routine. It ensuresthat the execution has reached the point where the checkpoint le being usedfor recovery was originally created, and that the application state has beencorrectly recovered.

Conditional sentences: Conditional blocks which contain one or more RECsneed to be tested, since not all processes might have executed the enclosedRECs in the original execution. For instance, consider an application thatperforms le I/O through one, and only one, of the parallel processes. I/Ooperations will be enclosed in a conditional expression checking the processID inside the parallel application. At restart time, conditional expressions areevaluated to ensure that each parallel process executes only RECs containedin appropriate conditional branches.

Stack recovery: Calls to procedures which recursively contain (i.e. throughany number of nested procedure calls) one or more RECs. This rebuilds theoriginal application stack.

REC detection is completely automated by the CPPC compiler, which is ableto automatically nd all described REC types. CPPC controls execution ow whenrestarting an application, making it jump from the end of an REC to the beginningof the next one, and skipping non-relevant code. The result is an ordered executionof selected state-recovering statements which, eventually, creates a replica of theoriginal application state.

1.7. Summary

This chapter describes the various alternatives for fault tolerance in parallelapplications in the literature. It is focused in MPI applications, since it is the

1.7 Summary 29

de-facto standard. We identied rollback-recovery as a better approach to faulttolerance than MPI extensions and modication of MPI semantics, for it preservesthe MPI interface and adapts better to modern heterogeneous environments. Thetwo core approaches to rollback-recovery, checkpoint-based and log-based, have beendescribed, along with three fundamental properties of checkpoint le creation: gran-ularity, transparency and portability. Then, existing checkpointing tools have beenanalyzed and categorized, outlining the drawbacks they present in the areas of scal-ability, eciency and portability. As long as each of them has its unique advan-tages, none manages to properly address all three problems at the same time. Fi-nally, CPPC (ComPiler for Portable Checkpointing), the checkpointing approachpresented in this work, has been described, outlining its most important charac-teristics and design decisions. It guarantees restart consistency through the use ofcompile-time coordination, a manner of implicit process coordination that does notrequire communications between processes to be performed at runtime. It providesa portable operation by employing a restart protocol that enables portable recoveryof non-portable objects through selective code re-execution.

From this point on, we describe the implementation of the two fundamentalparts of CPPC in greater depth: the checkpointing library, and the source-to-sourcecompiler.

Chapter 2

CPPC Library

This chapter discusses the design and implementation details for the CPPCcheckpointing library, which provides all necessary routines for checkpointing. It isimplemented in C++, and its design follows the model-view-controller (MVC) [16]design pattern in a way that allows the same core implementation to be used bydierent host programming languages. The model is structured in layers [16], decou-pling each functional level from the rest. The following sections provide an in-depthdescription of the library features, following a top-down approach that begins atthe application code (the view in the MVC design), describing what services the li-brary provides and how it provides them, while traversing deeper into the frameworkdesign, covering the controller and how it adapts external requests to the internalmodel and, nally, presenting the model and each of its layers.

2.1. View

In our MVC design, the view corresponds to the code of a parallel applicationthat uses CPPC in order to obtain fault tolerance. As in any MVC design, the viewis interchangeable so that the framework can service dierent contexts. For CPPC,this means allowing the same model to be accessed from dierent programming lan-guages. Specically, interfaces for C and Fortran 77 have been implemented, butspecifying interfaces for any language supporting calculated gotos1 is straightfor-ward. The need for calculated gotos is justied later on, and exists due to restart

1A calculated goto is a goto whose target is an array-indexed address.

31

32 Chapter 2. CPPC Library

protocol implementation details (see Section 2.1.7).

Since the number and semantics of functions exposed to the view is constant,they can be explained without referring to any specic implementation, except forillustrative purposes, in which cases we default to the C version of the interface. Thefollowing subsections describe the services CPPC provides to parallel applicationsthrough its API, briey summarized in Table 2.1. CPPC has two operation modes:checkpoint operation and restart operation. The checkpoint operation mode is usedwhen a normal execution is being run. It consists of marking relevant variables(variable registration) and saving them into state les at checkpoints in the code.Restart operation mode emerges when the application must be restarted from apreviously saved checkpoint le. Library functions will behave dierently dependingon the current execution state.

2.1.1. CPPC initialization and shutdown

CPPC needs initialization operations to be performed before servicing most ap-plication requests. The initialization operations must be spread amongst two dif-ferent function calls. The rst, CPPC_Init_configuration(), reads congurationparameters. Amongst others, it decides whether a restart is to take place. It shouldbe executed as soon as possible, so it can process and remove all CPPC command-line parameters that should not become visible to the application itself, to avoidsyntax errors. In its C version, this call receives the command-line parameters andtheir number, much like MPI_Init() does. After its execution, the library knowswhether or not to enable the restart protocol. If so, selective code re-executioneventually directs the execution ow towards the MPI_Init() call and, after it, tothe CPPC_Init_state() call. This second initialization function creates necessarymemory structures and, if a restart is taking place, calculates the recovery line andmakes restart data available to all parallel processes. Interprocess communicationsare required in order to nd the recovery line. Thus, the CPPC_Init_state() func-tion must be called after the MPI subsystem has been initialized. This is the reasonto spread the initialization operation into two dierent function calls.

As for the nalization of the library, the CPPC_Shutdown() function must becalled to ensure a consistent shutdown. Besides removing unnecessary les and free-ing memory, this call ensures that any multithreaded checkpoint operation nishesbefore allowing the program to exit.

2.1 View 33

Table2.1:

Interfacesummaryof

theCPPClibrary

Function

Checkpointoperation

Restartoperation

CPPC_Add_loop_index()

Noties

CPPCthat

theexecution

isaboutto

enteran

instrumentedloop

CPPC_Commit_call_image()

Storesparam

eter

values

Does

nothing

CPPC_Context_pop()

Noties

CPPCthat

afunctionhas

returned

CPPC_Context_push()

Noties

CPPCthat

afunctionisbeingcalled

CPPC_Create_call_image()

Noties

CPPCthat

anon-portablecallfollow

sCPPC_Do_checkpoint()

Dumpsstatele

Checksrestart

completion

CPPC_Init_configuration()

Initializesconguration

CPPC_Init_state()

Initializesstate

Initializesstate

Findsrecovery

line

Readscheckpointle

CPPC_Jump_next()

Returns0

Returns1

CPPC_Register()

Creates

new

register

Recoversdata

Recreates

register

CPPC_Register_descriptor()

Creates

new

ledescriptorregister

Recoversdescriptorstate

Recreates

descriptorregister

CPPC_Register_parameter()

Storesparam

eter

data

Recoversparam

eter

data

CPPC_Remove_loop_index()

Noties

CPPCthat

theexecution

has

exited

aninstrumentedloop

CPPC_Set_loop_index()

Noties

CPPCthat

theexecution

has

startedanew

iterationin

aninstrumentedloop

CPPC_Shutdown()

Waits

forcheckpoints

tocomplete

Not

used

Frees

mem

oryusedbyCPPC

CPPC_Unregister()

Deletes

existingregister

Not

used

CPPC_Unregister_descriptor()

Deletes

descriptorregister

Not

used

34 Chapter 2. CPPC Library

2.1.2. Variable registration

CPPC follows a variable level approach to checkpointing. This means that theapplication code must explicitly mark the variables that are to be stored in state les.We call this process variable registration. The function used to register a variableis CPPC_Register() which, in its C version, receives the following parameters:

Base address: The base memory address of the variable to be registered. ForC scalar variables, the address-of operator (&) should be used.

Number of elements: How many elements of type specied through the thirdparameter are being registered.

Data type: Since portability is one of the top CPPC priorities, stored dataneeds to be labelled in a way that allows for conversions to be performed ifnecessary. Besides, this labelling should be independent of the host program-ming language or the communications system used (MPI provides its own datatype codes). Thus, CPPC denes constant values associated to the basic datatypes to be registered. For instance, registration of character types is done us-ing CPPC_CHAR, CPPC_INT is used for integers, etc. When implementing viewsfor languages supporting overloaded methods the use of labels would not benecessary. Instead, a dierent function could be used for each data type.

Register ID: The base address is generally not unique for dierent variables(e.g. memory aliasing in C). Thus, each register is uniquely identied by astring. Any string that is unique for each function scope is valid. Therefore,the use of the variable name is recommended.

Memory type: A boolean value, indicating whether the memory to be reg-istered is static (scalar variables and arrays) or dynamic (pointers). This isrelevant when recovering the variable value, as will be further explained. Thisparameter is not present in languages such as Fortran 77 (which uses staticmemory only) or C++ (where partially instantiated [6] template functions [63]can be used to detect the dierence).

The register function behaves in a dierent way when the execution is beingrecovered from a state le. If called in this situation, it will not only perform theregistration, but also recover the original data found in the checkpoint le. Whenprocessing static variables, these data are copied to the memory allocated for them

2.1 View 35

by the compiler. For dynamic variables, the function returns a pointer to a memoryaddress containing the data. For instance, assume that an array containing 100integers named data is to be registered in a C application. The registration callsyntax would be:

CPPC_Register( data, 100, CPPC_INT, "data", CPPC_STATIC );

In a regular execution, the address of the data array is registered for futurestorage, containing 100 integers (the actual size in bytes depends on the computingplatform and is calculated internally). When the application is restarted, callingthis function recovers the original array values by copying them into the compiler-allocated memory. Also, as in a normal execution, a register for the variable iscreated. If data were not an array, but a pointer containing 100 integers, theregistration call would be:

data = CPPC_Register( data, 100, CPPC_INT, "data", CPPC_DYNAMIC );

Where the return value to the variable, and the memory type parameter indi-cates that this register involves dynamic memory. Upon application restart, the callreturns a memory address containing the original data, instead of replicating theminto the allocated memory.

A CPPC_Unregister() function is also provided for the removal of obsolete vari-able registers. Variables that have lost relevance are excluded from future check-points, thus optimizing state les sizes.

2.1.3. Non-portable calls

As seen in Section 1.6.2, CPPC uses selective code re-execution in order to recovernon-portable state. Function calls having a non-portable outcome are re-executedwhen recovering an application using the same parameter values as in the originalexecution. The CPPC library provides functions that store parameter values inorder to allow for proper re-execution. Before performing a non-portable invocation,a call to the CPPC_Create_call_image() function is issued. The parameters for thisfunction are its line number in the code and the name of the function itself. Thispair uniquely identies a specic non-portable call. Note that simply using the line

36 Chapter 2. CPPC Library

number as identier would not suce, since a line in the code may contain morethan one non-portable call (e.g. by using the return value of a non-portable call asa parameter for another non-portable call). Internally, CPPC creates a frame forthe parameters to be stored, named call image.

After creating the call image frame, parameters that are required to preservetheir values upon restart are passed to a special version of the registration func-tion, called CPPC_Register_parameter(), which receives the same parameters asCPPC_Register(). The dierence between both is that a registered parameter isassociated to a call image, rather than to a procedure scope. After registering allrequired parameters, the CPPC_Commit_call_image() function stores the call imagefor inclusion in subsequent checkpoint les. A typical example of use for instrument-ing an MPI_Comm_split() call would be:

CPPC_Create_call_image( "MPI_Comm_split", line_number );

CPPC_Register_parameter( &color, 1, CPPC_INT, "color",

CPPC_STATIC );

CPPC_Register_parameter( &key, 1, CPPC_INT, "key",

CPPC_STATIC );

CPPC_Commit_call_image();

MPI_Comm_split( comm, color, key, comm_out );

When restarting the application, the execution ow is directed towards this REC,and the values for color and key are recovered. Note that the original communicatorcomm is not registered, on account of it being a non-portable object. Either its valueis recovered by a previous execution of a non-portable call, or it is a MPI-denedconstant, such as MPI_COMM_WORLD. Eventually, the non-portable call is executedin the same conditions as in the original run, thus having the same semantic out-come: a new communicator comm_out, containing the processes in comm with thesame value of color, ranked attending to the values of key. Note that the specicMPI implementation used for restart could be dierent from the one used in theoriginal run, and the outcome would be semantically correct in the new executionenvironment.

The basic functional dierence between a regular variable register and a call im-age parameter is that in the former the variable address is saved and its contentsstored when the checkpoint function is called. The value of call image parameters,however, is stored in volatile memory when CPPC_Commit_call_image() is invoked,

2.1 View 37

and included in all subsequent checkpoint les. There is no way to remove a parame-ter registration, since a non-portable call never stops being relevant to the executionin CPPC's scheme.

2.1.4. Context management

The previous sections showed some of the data that CPPC includes in check-point les, namely variable registrations and call images. However, in order tocorrectly categorize each recorded piece of data into its correct scope, the CPPClibrary must be able to keep track of execution context changes. In order to doso, it internally manages a context hierarchy. Each context object represents acall to a procedure, and contains the information required for recovering data inthat procedure scope. Contexts contain variable registers, call images, and alsoother subcontexts, created by nested calls to the same or other procedures. Thishierarchical representation allows for the sequence of procedure calls made by theoriginal execution to be recreated upon restart. In this way, the application stackis rebuilt, and the relevant state is recovered inside its appropriate scope. This hi-erarchical representation allows for the application stack to be rebuilt upon restart,by recreating the sequence of procedure calls made by the original execution, whilerecovering the relevant state inside each procedure scope. This heap-shaped hierar-chy generalizes any pattern in procedure calling, including recursivity. Two CPPClibrary routines, CPPC_Context_push() and CPPC_Context_pop(), are used to no-tify context changes. These calls are only inserted before and after a call to aCPPC-instrumented procedure, respectively. A CPPC-instrumented procedure isany procedure containing CPPC code, be it registers, call images, checkpoints, orsimply calls to procedures containing any of these.

There is another situation that requires tracking by the library, and that is theinsertion of non-portable calls inside loops. If this happens, the non-portable callwill be repeated a certain number of times, and a dierent call image must becreated for each invocation. Moreover, CPPC must provide means to re-executethe non-portable section of the loop when recovery takes place. For this purpose,the CPPC_Add_loop_index(), CPPC_Set_loop_index() and CPPC_Remove_loop_-

index() functions are provided. Their purpose is to create a special type of context,called a loop context, which may only contain call images. Other types of registra-tion performed while inside the loop will be automatically inserted into the rstpredecessor node that represents a regular execution context. The following code is

38 Chapter 2. CPPC Library

a simple example of use of this type of context:

CPPC_Add_loop_index( "i", CPPC_INT );

for( i = 0; i < n; i++ )

CPPC_Set_loop_index( &i );

/* non-portable calls go here */

CPPC_Remove_loop_index();

Loop contexts need only be inserted on loops containing non-portable calls. Reg-ular variable registrations, checkpoints, or other CPPC operations potentially placedinside a loop do not need any special modications to their environment.

2.1.5. Open les

Files are characterized as being random access dataows, with a state denednot only by their data or path, but also by a pointer to the current position insidethe data stream. This position must be recovered when reopening takes place. TheCPPC_Register_descriptor() function is used to track open les. It receives aunique identier for the descriptor (an integer), the base address for the descriptoritself, the type of the descriptor being passed, and the path of the le being opened.An example of use:

fd = open( path, mode );

CPPC_Register_descriptor( id_number, &fd, CPPC_UNIX_FD, path );

During a regular execution, the register descriptor function marks the le asopen, so that relevant information about the descriptor, including the position ofthe stream pointer, will be stored when a checkpoint is reached. When recoveringthe application, the library re-executes the open function and then the register func-tion, which in this context ensures that the stream pointer is correctly repositioned.The CPPC_UNIX_FD parameter tells the library that a regular UNIX le descriptor,represented as an integer, is being registered. Another valid value for the descriptortype is CPPC_UNIX_FILE, used for FILE * descriptors. More register types can beadded to the library by implementing the appropriate handlers and functionalitiesinside the lesystem abstraction module (see Section 2.3.4).

2.1 View 39

A CPPC_Unregister_descriptor() function is also provided to remove descrip-tors after their associated le has been closed, eectively removing any further ref-erence to it on future checkpoint les.

2.1.6. Checkpoint le dumping

As mentioned in the previous chapter, calls to the checkpoint function, CPPC_-Do_checkpoint(), must be introduced at safe points, where no in-transit nor incon-sistent messages between processes exist. From the application's perspective, thiscall simply receives a unique integer parameter which is used to identify the pointin the code where the state dumping is performed, so that it is distinguishable fromother checkpoints performed in the same execution context, allowing the correctidentication of the restart point during the recovery process.

2.1.7. Application restart

The restart phase has three fundamental parts: nding the recovery line, read-ing the checkpoint data into memory, and recovering the application state. Therst two steps are encapsulated inside the CPPC_Init_state() call. However, theactual reconstruction of the application state and repositioning of control ow mustbe reached through the ordered execution of the RECs in the application (see Sec-tion 1.6.2). For this purpose, jump labels are inserted into the appropriate positionsin the code, and a structure indexing these labels is dened. In the C interface, thisstructure is an array of ordered jump addresses (void * values) for the dierentlabels. The restart process consists of jumping from the end of an REC block to thebeginning of the next one, executing its code and jumping again. This process isrepeated until the library determines that the restart has nished and that a regularsequential execution of the code can be enabled. The C function the applicationuses to query for the restart state is CPPC_Jump_next(). It returns 0 if no restartis being performed, and a dierent value otherwise.

Note that jumps must be performed inside the application code without leavingthe current procedure scope. Otherwise, the call stack would be inconsistent for thetarget scope. Also note that there is no more elegant solution than the use of gotos,since RECs can be found at any point or nesting level inside the code, discarding theuse of switch constructs. In this situation, the use of arrays indexing local jump

40 Chapter 2. CPPC Library

addresses is a safe and coherent solution, as long as it is guaranteed that said arraysare not passed as parameters between functions, which CPPC does not do.

Jumps used to reach a REC from the end of another one are called conditional

jumps, since they are taken inside an if structure controlled by CPPC_Jump_next().Their code is as follows:

if( CPPC_Jump_next() )

int jump_counter = cppc_next_label++;

goto *cppc_labels[ jump_counter ];

Jump labels and conditional jumps are inserted enclosing all the six types ofRECs: library initialization, portable state recovery (variable registration), non-portable state recovery (execution of calls with non-portable outcome), restart com-pletion tests (done by the checkpoint function), conditional sentences and stack re-covery (execution of calls to CPPC-instrumented functions). For some of these RECtypes, however, the control ow instrumentation varies slightly. The CPPC_Init_-

configuration() call starts the jump sequence and is usually the rst statementin the code, and therefore does not include a jump label before it. Also, conditionalsentences and stack recovery require a more complex version of the control owinstrumentation, as detailed in the following paragraphs.

Calls to CPPC-instrumented procedures must be re-executed during restart. Forthis purpose, CPPC_Context_push() and CPPC_Context_pop() are placed enclosingthe call, which is itself executed as well. Internally, control ow code is added aspreviously explained, but with the rst statement in the procedure being a condi-tional jump towards the rst REC inside its code, and with the last conditional jumpreaching a generic return instruction. Note that the generic return does not aectthe state recovery process. If the returned value is stored into a relevant variable,then this variable will be registered after the call, and its correct value recovered.Otherwise, the returned value is discarded or stored into a non-relevant variable. Inboth cases, the generic return cannot aect the execution outcome.

Adding control ow code to an instrumented conditional sentence is similar todoing so to instrumented procedures, just more complex. The conditional expressionmust be re-evaluated to ensure that each parallel process executes the proper branch.Variables involved in the computation of the control expression are registered as if

2.2 Controllers 41

they were used in a non-portable call, using the call image-related functions previ-ously detailed. Thus, it is guaranteed that the evaluation of the control expressionis consistent with the original one. Each conditional branch is then modied so thatits rst statement is a conditional jump towards the rst REC in that branch. Atthe end of each branch, a conditional jump is placed that directs the process to thenext REC outside the conditional construct. These particular conditional jumpsare dierent from the regular ones in that they do not increase the jump counterby one, but by an amount dependent on the number of RECs between the currentposition and the desired target. For instance, the rst conditional jump in the thenbranch of an if statement will increase the jump counter by one. However, the lastconditional jump in that branch will increase the counter by the number of RECscontained in the else branch plus one. Similarly, the rst jump in the else branchincreases the counter by the number of RECs contained in the then branch plusone, while the last one increases the counter by one. This example can be easilygeneralized for other conditional constructs, such as switch.

The presented restart protocol fullls the portability objective in CPPC. It doesnot directly modify non-portable state such as the program counter or the appli-cation stack, but instead uses the application code itself to perform required mod-ications using workarounds. It does not impose non-portable constraints such asrecovering the variables in the same memory addresses as in the original run. It alsois implementable on a wide range of programming languages. Although requiredcode modications may seem cumbersome and nontrivial, they are all automaticallyperformed by the CPPC compiler described in the next chapter.

2.2. Controllers

To connect the dierent views with the library model, CPPC's API is imple-mented into controllers that abstract the particularities of the chosen programminglanguage and translate the requests to the set of C++ functions implemented by themodel. Each controller performs dierent actions depending on its target language.This section covers the controller connecting the model with C views. Given thesimilarities between this language and C++, the controller is extremely simple. Thethree fundamental remarks about its implementation are:

For connecting C and C++, the taken approach is to implement the controllerin C++ declaring their functions as extern "C". Thus, the compiler generates

42 Chapter 2. CPPC Library

C function signatures, which may be linked with the parallel application.

Inside the functions declared as extern "C" the controller interacts with themodel façade. Said façade is a singleton2 that exposes a series of operationsalmost identical to those seen in Table 2.1. The controller perfoms typeconversions, such as converting C-style strings (char *) to the C++-stylestd::string.

Since repositioning a le stream pointer is an operation which depends onthe host programming language, it cannot be implemented inside the model.Instead, controllers are responsible for this task. The C version of the controllerkeeps track of the set of registered le descriptors and, upon restart, getsthe appropriate le oset from the model and repositions the pointer streamusing the lesystem abstraction interface found in the portability layer (seeSection 2.3.4).

2.3. Model

The core implementation of the CPPC library functionality is found inside themodel. Figure 2.1 oers an overview of the framework design. The model is struc-tured in layers which decouple each functional level from the rest. It oers a façadewhich serves as centralized access point, implementing one function for each of theones exposed in the interface. Under it there is a checkpointing layer, which receivesa checkpoint le in a specic format and implements protocol-dependent operations.Then it passes the data to the writing layer, which is an interchangeable piece ofsoftware in charge of writing and reading checkpoint les to and from stable stor-age. There are two additional layers: a portability layer, which abstracts operationswhich depend on operating system services; and a utility layer, which provides im-plementations of useful operations, such as data compression or integrity checks,which are commonly used by the other layers. All these are described in subsequentsubsections. Note that, at this level, the library code is completely independent ofthe controller (C, Fortran 77, etc.).

2In object-oriented programming, a singleton is an object programmed using a design patternthat characterizes a class by exposing a single access point to its instances. This access point mustguarantee global accessibility and the existence, at any given time, of a single instance of the class,which is returned in each invocation. There are many variations of this pattern depending on thegoals to be achieved. For more information see [6, 28].

2.3 Model 43

Figure 2.1: Global design of the CPPC framework

2.3.1. Façade

This layer orchestrates all the model behavior, while oering an external interfaceto the controller. It implements the following functions:

• CPPC_Init_configuration()

This function initializes the library conguration. CPPC implements a cong-uration manager that centralizes the access to dierent conguration sources intoa single interface that abstracts this functionality. Due to the modular design ofCPPC, the naming convention for conguration parameters is hierarchical, eachlevel referring to a specic module or layer. How parameter names are constructeddepends on the source of each of them, but the rst level is always CPPC. The con-guration manager adopts the following rules for parameter naming:

1. Command-line parameters: hierarchical levels are separated using the / char-

44 Chapter 2. CPPC Library

acter. For instance, the parameter controlling whether a restart is to take placeis named CPPC/Facade/Restart.

2. Environment parameters: Environment variables cannot contain the / char-acter. Thus, _ is used for separating hierarchy levels. The restart parameteras dened by an environment variable is called CPPC_Facade_Restart.

3. Conguration les: A convenient way to specify parameters which remain con-stant through many executions. A simple DTD3 has been created to describean XML scheme for conguration les. There are basically two types of tags:variables, which have a name and a value; and modules, which may containvariables and other modules. A simple example of conguration le deningthe restart parameter would be:

<cppc>

<module name="Facade">

<variable name="Restart" value="false"/>

</module

</cppc>

The library is also able to accept conguration les in plain format in case noXML parser is available in the computing platform. In this case, the namingconvention is the same as that of command-line parameters, and a simple lecontaining the restart parameter would look like:

CPPC/Facade/Restart=false

Some common conguration parameters are described in Table 2.2. However,since congurable plugins having their own conguration options may be dynami-cally added to the CPPC library, this is not a comprehensive list.

• CPPC_Init_state()

Once CPPC has been congured, it can initialize its working environment. First,it creates a directory for state le storage. To support shared le systems (common-place in clusters), a directory tree is created:

3A DTD, or Document Type Denition, is an XML schema language primarily used for theexpression of a schema via a set of declarations that describe a class, or type of document, in termsof constraints on the structure of that document.

2.3 Model 45

Table 2.2: CPPC conguration parametersModule Parameter Description

Facade

RootDirDirectory where CPPC will store andfetch checkpoint les.

ApplicationNameSubdirectory of RootDir wherecheckpoint les will be eectivelystored.

Restart

Controls whether the applicationmust be restarted. It is recommendedto set this to false and enforcerestart using command-line parame-ters.

FrequencyNumber of calls to CPPC_Do_-

checkpoint() between eectivestate dumping.

CheckpointOnFirstTouch

If set to true, the state dump-ing will be performed every rsttime that a particular call toCPPC_Do_checkpoint() is reached.

Sux Sux for the generated les.

StoredCheckpointsMaximum number of checkpoint lesstored per node before beginning dele-tion of older ones.

DeleteCheckpointsIf set to true, stored checkpointswill be removed upon successful ex-ecution.

Compressor TypeType of compressor used. See Sec-tion 2.3.5.

Writer TypeType of writer used. See Sec-tion 2.3.3.

Checkpointer MultithreadedUse multithreaded dumping. See Sec-tion 2.3.2.

46 Chapter 2. CPPC Library

Level 1. One folder to store all CPPC applications data. E.g. /tmp/CPPC_Files.

Level 2. Inside the rst level, a folder for each application. The name ofthis folder is determined by the application identier dened through theCPPC/Facade/ApplicationName parameter seen in Table 2.2. For instance,an application named DBEM would nd its les under /tmp/CPPC_Files/DBEM.Automatically determining the application name using the rst command-lineparameter was discarded, since:

• Executing an application on queues makes the application name variabledepending on its PID4. Thus, an application being re-executed with adierent PID will not nd old checkpoint les created during a regularexecution. Besides, if the aforementioned problem did not exist, twodierent executions of the same application would try to share the samefolder name, leading to consistency problems.

• Not all programming languages accept comand-line parameters. Fortran77, for instance, does not dene a standard for accessing them (althoughmost compilers provide their particular way to do so).

Level 3. Finally, a folder per process. Each one of them may obtain a uniqueidentier in the context of the parallel application execution as their rank5 inthe MPI_COMM_WORLD communicator. Thus, a process with rank 4 executing aCPPC application would store its les under /tmp/CPPC_Files/DBEM/4/.

If it is not necessary to restart the application from a previously created check-point (depending on the CPPC/Facade/Restart parameter), the initialization pro-cess has nished, and the control ow returns to the parallel application. Otherwise,a valid recovery line has to be found.

To determine which les should be selected for restarting, each process mustcheck which checkpoint les are available to it, and communicate with the rest to ndthe most recent le common to all processes. Since it is guaranteed that all processestake the same number of checkpoints, les with the same name6 in dierent processescorrespond to the same point in the execution of the application. Since no in-transitnor inconsistent messages existed upon creation, the set of homonym checkpoint les

4PID, or process identier, is a number used by some operating system kernels (such as that ofUNIX, Mac OS X and Windows NT) to uniquely identify a process.

5MPI assigns each process an integer which uniquely identifes it in a communicator.

2.3 Model 47

A = Available state files, ordered by

unique file code

agreement ← false

While not agreement

N ← Newest correct state file in A

If all processes propose N then

agreement ← true

Else

O ← Older restart point proposed

NO ← x ∈ A / x is newer than O

A ← A-NO

End If

End While

Delete files older than N

Figure 2.2: Pseudocode of the algorithm for nding the recovery line

forms a strongly consistent global state. A pseudocode of the algorithm for ndinga valid recovery line is shown in Figure 2.2.

Once the recovery line has been found and each process knows which le itmust use for restart, it reads its contents into memory. These data are ready to berecovered as the application progresses through its RECs.

• CPPC_Shutdown()

This function nalizes the library, leaving it in a consistent state. First, it ensuresthat the model being shut down was previously initialized through calls to the ini-tialization functions. Then, if that rst check succeeds, it noties the inferior check-pointing layer that a shutdown is taking place, so that it may take appropriate ac-tion. This may involve waiting for any multithreaded checkpoint operations to nishbefore proceeding. After that, it evaluates the CPPC/Facade/DeleteCheckpoints

conguration parameter. If set to true, it removes all generated checkpoint les, andbegins traversing the directory hierarchy built by CPPC_Init_configuration(), re-moving all empty directories, until it nds a non-empty one or removes the hierarchy

6CPPC assigns le names incrementally, beginning at 0, and increasing for each newly createdcheckpoint le.

48 Chapter 2. CPPC Library

root. Since the lesystem interface is not standard in C++ (classes for le access areprovided, but not for folder manipulation), CPPC includes an abstraction layer forlesystem access to facilitate the portability of the library. The implementation ofthis layer is operating system-dependent. For more details, see Section 2.3.4. Afterthese steps have been taken, the façade instance is destroyed, library memory freed,and the application is ready to exit.

• CPPC_Register()

This function is in charge of variable registration: scheduling memory regionsto be stored in subsequent checkpoints. This function behaves in a dierent way ifcalled during a regular execution or during recovery.

During a regular execution, the façade is responsible for managing the set ofregisters performed by the application. Thus, when this function is called, a newregister is added to this set as a structure containing its base memory address; theend memory address; the data type, which may be necessary in order for inferiorlayers to correctly store it; and the register name. If the name conicts with anotherone in this execution context, the register will not be created. For this reason, aspreviously stated, the use of the variable name is recommended (no two variableswith the same name may exist in the same procedure scope).

In order to calculate the end memory address from the base address, the datatype and the number of elements of the register (which are the parameters providedto the function), it uses another of the services provided by the portability layer, inthis case a data type manager, that through partial template instantiation is ableto provide a single C++ implementation to functionalities such as obtaining thesize of a data type, independently of the architecture on which the library has beencompiled.

If the application is being restarted, however, it is necessary to recover the dataoriginally stored in the registered memory and restore them to the place where theapplication expects to nd them. Since the checkpoint le contents were recoveredwhen CPPC_Init_state() was called, the only thing to be done at this point is tomake said contents reachable for the application. If the variable being registered isstatic, the library copies the stored data to the base address provided. Otherwise,it returns a pointer to the appropriate memory address, that the application mayassign to the registered pointer variable. After data recovery is nished, the library

2.3 Model 49

recreates the old register as a regular execution of this call would.

• CPPC_Unregister()

This is one of the simplest functions in the façade. It receives a register nameand removes the corresponding match from the set of existing registers. Note thatthe memory address does not uniquely identify a register, since aliasing betweenvariables may cause dierent registers to start at the same memory address.

• CPPC_Create_call_image()

This function call creates a new call image object for inserting required parameterregistrations. During a recovery, the function also fetches the old call image from thestate le in order to subsequently recover the aforementioned parameters' values.

• CPPC_Register_parameter()

At the façade level, this function does not dier much from the regular variableregistration. The only dierence is that the register is inserted inside the object cre-ated by the last CPPC_Create_call_image() function, instead of being associatedto the current execution context. During restart, the function reads the registerfrom the recovered call image and recovers the parameter value.

• CPPC_Commit_call_image()

When a normal execution is run, the commit function performs aliasing checks(see the CPPC_Do_checkpoint() section) on the registered parameters inside theassociated call image and stores their values in volatile memory to be stored atsubsequent checkpoints. During restart, it simply removes the old call image object,which should be now empty after all contained parameter registrations have beenrecovered.

50 Chapter 2. CPPC Library

• CPPC_Context_push()

Called when the execution is going to perform a call to a function containingCPPC instrumentation. During a normal execution, it creates a new executioncontext object and adds it to the context hierarchy in the façade under the currentone. After that, it assigns the current execution context pointer to reference thenewly created one. During restart, the function handles two dierent homomorphichierarchies: besides the regular one, the saved context hierarchy is recovered fromthe checkpoint le and traversed as the execution progresses. The façade maintainstwo dierent pointers to current execution contexts: one for the current executionhierarchy, and another one for the one recovered from the state le, which containsall registers, call images, and checkpoints taken during the original execution.

• CPPC_Context_pop()

Called when the execution has returned from a call to a CPPC-instrumentedfunction, it updates the execution context pointer to reference the parent of the ex-isting one. Additionally, if the context the execution is leaving is empty (it containsno registers, checkpoints, call images, or subcontexts), it destroys the object, so thatit is not saved to subsequent checkpoints. During a restart, this call also updatesthe pointer to the current context in the hierarchy recovered from the state le.

• CPPC_Register_descriptor()

This function adds an entry to the set of open les to be recorded at subsequentcheckpoint les. It does not perform any kind of stream pointer positioning atrestart, which is the responsibility of the view as previously explained.

• CPPC_Unregister_descriptor()

Removes an entry from the set of current open les.

• CPPC_Do_checkpoint()

This is another function with a behavior that depends on whether a normalexecution or a restart is being done. In the rst case, its purpose is to initiate the

2.3 Model 51

state le creation. There are some conguration parameters that aect the outcomeof this operation. First, it is possible to dene a checkpointing frequency. It isspecied as the number of calls to the CPPC_Do_checkpoint() function before anactual checkpoint takes place. If CPPC/Facade/Frequency equals N, then the statele dumping will only be performed each N calls to this function. This parameter isintended to be used in computational loops containing checkpoint calls.

Alternatively, it is possible to dene another parameter that forces checkpointsto be taken the rst time the execution reaches them. This option is included toavoid skipping checkpoints outside computational loops. These are usually placed atinection points in parallel programs, where a task has been nished and a new oneis going to begin. Usually, after a computation phase is nished its partial resultsare available to be stored while auxiliar variables used for intermediate results arenot relevant anymore, thus reducing state le sizes.

Summarizing, a checkpoint will be taken if: (a) the number of calls is a multipleof the frequency parameter; or (b) it is the rst time that a particular checkpoint isreached in the code.

After deciding that a checkpoint le has to be created, the rst thing the façadedoes is to perform aliasing checks on all existing registers. It does so by traversingthe context hierarchy structure, starting at its root (the context associated to themain procedure), and analyzing all registers in each context. During this operation,the function creates a list of memory blocks to be stored in the checkpoint le.For each register, it analyzes its beginning and end memory addresses and comparesthem to the addresses of other registered variables. If no overlap exists, it schedules anew memory block for inclusion in the checkpoint le, corresponding to the memoryassigned to the analyzed register. Otherwise the smallest memory block containingall overlapping registers' memory is calculated. Thus, a single memory block inthe checkpoint le may be assigned to multiple registered variables. The resultof this calculation is the basic checkpoint le format depicted in Figure 2.3. Ascan be seen, each state le is divided in two dierent parts: a metadata sectionand an application data section. The application data section contains the memoryblocks copied from the application memory, which hold the relevant data necessaryfor restart. The metadata section contains information for the correct recoveryof the application memory: the execution context hierarchy, containing variableand parameter registrations. The gure omits call images for simplicity, since theyare similar to execution contexts, but containing parameter registrations only. Foreach variable to be recovered, execution contexts store information about its type,

52 Chapter 2. CPPC Library

Figure 2.3: Data hierarchy format used for writing plugins

associated memory block in the application data section, and osets to calculatethe beginning and end of its data inside that memory block. The use of portableosets instead of memory addresses [64] enables pointer portability. The osets arecalculated taking into account the original address for the registered variable and theaddress for the corresponding block in the application data section. Upon recovery,the memory block will be loaded into memory, and memory addresses for variables tobe recovered will be back-calculated using the stored osets. This preserves aliasingrelationships through application restarts.

Note that, while CPPC provides a mechanism for dynamically plugging newwriting modules, which perform the data writing to stable storage (see Section 2.3.3),all of them must accept this checkpoint le format as input.

If an execution is being restarted, this function does not perform checkpointing.Rather, the library tests two conditions: (a) that this is the execution context wherethe checkpoint le used for restart was created; and (b) that this is the exact pointinside the procedure code where it happened. If both are fullled, the restart isover, conditional jumps are deactivated, and the parallel application will be allowed

2.3 Model 53

to normally resume its execution.

• CPPC_Jump_next()

As mentioned when discussing the library views, the purpose of this function isto notify the application as to whether a restart or a normal execution is takingplace. It returns a boolean value indicating the current state. The application usesthis return value to control conditional jumps.

2.3.2. Checkpointing layer

The façade implements most of its functionality through dierent sublayers. Pre-cisely, it uses the checkpointing layer and some of the functionalities oered by theportability layer.

The checkpointing layer provides access points for managing the creation of thecheckpoint itself. The façade handles, basically, execution contexts, call images,and variable and parameter registrations. When the time comes to include thosein a checkpoint le, it uses the services provided by the checkpointing layer, whichencapsulates the particularities of the checkpointing protocol itself. This allows forprotocol changes without aecting the façade, nor the layers below, either. The ser-vices provided by the checkpointing layer are described in the following subsections.

• State le writing

The fundamental goal of this service is to implement the specics of the check-point protocol before delegating the le creation into the layer below. If a coordi-nated checkpointing scheme were to be introduced, this would be the layer to do so.Since CPPC's compile-time coordination checkpointing scheme does not use runtimecoordination or message-logging, there are no relevant actions to be taken at thispoint.

This service checks that the data passed by the façade are valid, creating a newthread to handle state dumping (multithreaded checkpointing) and return. How-ever, it also needs to ensure that the two threads do not incur inconsistent memorymodications, since accessing the checkpoint data is a critical section. In order toavoid this situation, the checkpointing layer creates a replica of the memory blocks

54 Chapter 2. CPPC Library

to be dumped before creating the checkpointing thread. This enables concurrency,but it takes time to perform the copying itself. Another option would be to protectthe memory to be dumped, allowing only read accesses to it. This can be done byany modern operating system, but it presents several problems that make memoryreplication a better option:

There is no standard interface for memory protection, which makes the codenon-portable. This disadvantage can be easily overcome, however, since mem-ory protection could be abstracted through the portability layer.

In most systems the memory protection works at page level. Since CPPCworks at variable level, this scheme is too coarse-grained, which would resultin many variables being unnecessarily protected when located in the same pagethan relevant variables, resulting in a potential eciency loss.

When a process tries to write to protected memory, it receives a SIGSEGV. Inorder to avoid execution failure, the library would need to capture and handlethis signal while checkpointing takes place. This could mask a legitimatesegmentation fault caused by the application code for reasons dierent thana write to protected memory. This, however, should not happen in correctcodes.

Once memory replication is nished, a new thread is created to handle check-pointing le write, performed by the writing layer. CPPC supports sequential check-pointing through the CPPC/Checkpointer/Multithreaded conguration parameter.If set to false, no threads are created and the checkpointing process uses the orig-inal application memory, returning control when the state le has been persistentlycreated in stable storage.

This service is the only one from the checkpointing layer that the façade usesduring a normal execution. The rest of them are only relevant during the restartphase.

• Integrity checks

When recovering an execution, each process checks which local checkpoints it hasavailable to build the recovery line. Since CPPC aims to provide fault tolerance, it

2.3 Model 55

cannot overlook the fact that checkpoint les may become corrupt due to a series offacts. Therefore it must perform state le integrity checks.

CPPC provides an abstraction layer on the writing algorithm used for checkpointle creation. This is due to the desire to provide exibility and support for dierentcomputing platforms. However, this also means that integrity checks must be per-formed by the writing module that created a certain state le. The implementationof this service, therefore, only checks that the requested le exists, that it is notempty, and delegates all further actions to the writing layer.

• State le reading

This operation is also delegated to the writing layer. The specic writing moduleused for creating a certain checkpoint le is in charge of recovering its data backfor application restart. The only constraint on physical checkpoint le formats isthat their rst byte must uniquely identify the specic writing module used for itscreation. Thus, at this point, the checkpointing layer reads that rst byte, andobtains a writing module object by requesting an abstract factory [28] to provide awriter instance identied by that byte. The read operation is then delegated to theobtained writer, that returns the checkpoint in the hierarchical format previouslyseen in Figure 2.3. This checkpoint object is stored by the checkpointer for servicingfuture recovery requests by the façade.

• Partial data recovery

This operation is invoked after a checkpoint le has been selected and readinto memory. The checkpointing layer is in possession of all checkpoint data, andmust return memory blocks to the façade on demand. A partial data recoveryoperation is invoked to recover the contents of a variable or parameter registration.The checkpointing layer fetches the requested register from the current executioncontext in the saved hierarchy, and returns the address containing its data.

One of the parameters passed to this function is the size of the register to berecovered, in bytes. This number must be the same as the size of the matchingregister recovered from the checkpoint le. Note that this number depends on thebinary representation for the registered data type in the computing platform usedfor restart. This mechanism is used as an addition to le integrity checks to furtherguarantee consistency. The typical errors detected by this technique are inconsis-

56 Chapter 2. CPPC Library

tencies between the application code and the le being used for restarting, suchas situations where the code has been modied and recompiled after the le wasgenerated.

• Restart completion check

Besides storing variable registrations, call images and subcontexts, an executioncontext inside a checkpoint le stores whether that le was created while inside saidcontext or not. This function checks two conditions:

The current execution context in the saved hierarchy is marked as the onewhere the checkpoint le was created.

The checkpoint call in the current procedure is the one where the checkpointle was created. This is done through the unique identier passed to the check-point function, and ensures consistent restart when two or more checkpointfunctions are placed in the same procedure scope.

2.3.3. Writing layer

This layer is in charge of actually handling the stable storage of state les. It isabstracted through a pure virtual class [63] that provides four operations: write acheckpoint le, read a checkpoint le, check the integrity of a checkpoint le, andobtain a writer plugin identier code. This code is unique to each implementationof the virtual class, and guarantees a coherent management of the available writingstrategies. CPPC includes an HDF-57 [27] writing plugin. In the past, a binarywriter was also provided, which intended to provide eciency in exchange for func-tionality. It was discontinued, however, since conguring the HDF-5 library to writedata in memory format while tagging it to support data conversions when restartingthe application, if necessary, resulted in an equally ecient operation.

New writing plugins can be easily implemented and integrated into the library. Awriter abstract factory provides a function for writing plugins to register themselveswhen the application starts. The factory then manages a hash table indexed bywriter codes, and containing a constructor function for associated writer objects.

7HDF-5 is a hierarchical data format and associated library for the portable transfer of graphicaland numerical data between computers.

2.3 Model 57

Plugins are dynamically registered taking advantage of the initialization of constantvariables [6]. The factory provides a method for obtaining a writer through its writercode. Thus, the checkpointing layer may perform requests for writer instances intwo dierent ways, depending on the execution state:

If a state le is being written, the writer code is located through the CPPC/-

Writer/Type conguration parameter.

If a state le is being read, the writer code is obtained by reading the rstbyte of the checkpoint le used for restart.

Again, the only constraint for the physical format of generated state les is thattheir rst byte must contain the writer code of their creator. Otherwise, the writingstrategy must only guarantee that it is able to consistently recover its written data.

Regarding writing options and functionalities, these are completely dependenton the specic writing module being used. Each particular writing module can useits own namespace to accept conguration parameters (e.g. CPPC/Writer/HDF-5 forthe HDF-5 writer).

2.3.4. Portability layer

A portable tool for checkpointing parallel applications must deal with non-standard interfaces that may risk portability of the library between dierent hard-ware/software platforms. For instance, operating systems provide dierent inter-faces for accessing the lesystem. CPPC abstracts certain parts of its functionalitythrough a system of template classes [63] that receive a policy [6], or implementationclass in which they delegate all their behavior.

The portability layer is static, as oppossed to dynamic ones such as the writinglayer. This means that the actual implementation for the abstracted interface isselected at compile time. A recompilation is needed in order to change it later.Since this layer tries to achieve portability, and not extra functionalities, there is noneed to make it dynamic, which would in turn diminish eciency (by forcing theuse of virtual functions [63]).

58 Chapter 2. CPPC Library

• Filesystem

The problem of portably accessing the lesystem has been mentioned severaltimes. To workaround it, CPPC denes two interfaces that must be implementedfor each operating system to be supported by the library. The rst interface is thelesystem manager itself, providing operations for creation and removal of folders,removal of regular les (there are standard means for creating regular les in C++),test the existence and size of les, and for opening directory streams. Here the secondinterface comes into play: the directory stream itself. It is a pure virtual class thatmust provide a single method to iterate through the directory's contents. Bothinterfaces provide the necessary abstractions for CPPC operations to be portablyimplemented in the above layers.

A POSIX8 implementation of these interfaces is provided with CPPC. It maybe used for compiling the library in operating systems conforming to the standard.However, the implementation of these interfaces is quite simple for any given oper-ating system. Currently, a proposal of the Boost.Filesystem library9, which providesportable facilities to query and manipulate paths, les and directories, has been ac-cepted by the C++ Standards Committee for inclusion in the C++ Technical Report2. If nally included into the language, this would eliminate the need for this ab-straction layer. The inclusion of a Boost.Filesystem dependency in CPPC was alsoconsidered, but ultimately discarded since the Boost libraries are not commonplace.

• XML manipulation

This simple abstraction interface decouples CPPC from the specic XML parsingsolution being used. It only contains two relevant functions: one for indicating thele to be parsed and another one for recovering the value of a given tag.

An implementation of this interface using the Xerces-C [72] library, a portablewidespread parsing facility, is included with CPPC.

8POSIX (Portable Operating System Interface) is the collective name of a family of relatedstandards specied by the IEEE to dene the API, along with shell and utilities interfaces forsoftware compatible with variants of the Unix operating system, although the standard can applyto any operating system.

9The Boost C++ libraries are a collection of peer-reviewed, open source libraries that extendthe functionality of C++. They cover a wide range of application domains, ranging from general-purpose libraries (e.g. smart pointers implementations) to OS abstractions like Boost.FileSystem.The Boost libraries homepage can be found at http://www.boost.org.

2.3 Model 59

• Communications system

While CPPC was designed to work with MPI applications, the portability of thestrategies used for rollback-recovery allows it to be independent from the commu-nications system. Thus, it is enough to abstract communication-related operationsthrough an interface to achieve independence. Operations included in this interfaceare: global synchronizations (barriers); obtaining a unique rank for each processexecuting the application; obtaining the total number of processes executing theapplication; and performing reduction operations on the minimum and maximumof a basic data type. Note that if the communications system does not supportreduction operations, these might be implemented using global or point-to-pointcommunications.

Two dierent implementations of the communications interface are provided: anMPI-based one, and another one for the execution of sequential applications, calledno-communicator, that makes a trivial implementation of these operations in asingle-process environment.

2.3.5. Utility layer

This layer includes reusable functionalities useful for implementing writers, com-munication modules, etc.

• Compression system

Under certain circumstances, the user may need to apply compression to thestate les generated by CPPC. Originally, le compression was centralized by thecheckpointing layer. However, the abstraction between this and the writing layergenerated troublesome situations. For example, compressing a le involved writingit to disk uncompressed through the writing layer, reopening it at the checkpointinglayer, compressing the data and writing the le again. This becomes extremelyinconvenient when stable storage is remotely provided. Due to this, the systembecame an independent entity as part of the utility layer. Each writing plugin maydecide how to use (or not use) the compression algorithms included with CPPC. Thisenables the use of ad-hoc compression systems when using complex data storagelibraries for checkpoint le storage, like HDF-5, which provides its own compressioncapabilities. The obvious disadvantage is that decentralizing this operation increases

60 Chapter 2. CPPC Library

the complexity of the writing modules, which have to include the compression codethemselves.

The structure of the compression system is a replica of the writing layer. Com-pression algorithms must implement a common interface, which provides meth-ods for data compression and decompression, and register themselves with an ab-stract factory when the execution starts. Each plugin implementation has its ownunique identier code, so that the factory knows which instance to return. TheCPPC/Compressor/Type conguration parameter denes which implementation ofthe compressor to use. Writing plugins using the compression system must some-how store the code for the used compressor into the le, in a similar way to thewriter code, in order to enable decompression upon restart.

The impact of compressing les on checkpointing overhead is analyzed in Chap-ter 4.

• Conguration manager

This subsystem was introduced when covering the CPPC_Init_configuration()function in the façade, in Section 2.3.1. Its purpose is to encapsulate congurationsources and priorities. It consists of a singleton class with two public methods: oneto initialize the manager, which receives the application command-line parametersand initializes and reads CPPC's conguration le; and another one for queriesregarding parameter values.

Since the manager is heavily used by all the layers in the model, it uses a pa-rameter cache to improve the performance of the query operations.

• Cyclic redundancy checks

This system includes a reusable class implementing a common CRC-32 algorithm.It may be used by writing plugins to implement their integrity checks.

2.4. Summary

This chapter describes one of the two fundamental parts of this work: the check-pointing library for message-passing parallel applications. The design and imple-

2.4 Summary 61

mentation of the library have been described using a top-down approach, beginningat the application code and descending through the controller, used for decouplingthe model from the programming language being used, and nally through the modellayers of CPPC: façade, checkpointing layer and writing layer. Also, the portabilitylayer has been described, being specially important for providing abstractions forcertain OS-dependent operations (lesystem access, communications, etc.). Finally,the utility layer, which combines useful functionalities oered to all the model layers,was described.

As seen throughout the chapter, the modications that need to be performedon the application's code in order to achieve integration with CPPC are far fromtrivial. To automate this task, a source-to-source compiler is used, as described inthe following chapter.

Chapter 3

CPPC Compiler

The integration of the CPPC library with a parallel application requires modi-cations to be performed to the original code. Their specics were detailed in theprevious chapter. If manually performed, the process would place an undesirableburden upon the users, who would have to manually perform complex analyses suchas detecting live variables and safe points. Besides inserting CPPC library callsinto the appropriate places, users would also need to insert the control ow codethat enables the recovery process. In order to free users from these tasks, fulllingthe transparency goal in CPPC, a source-to-source compiler has been implementedthat automates all necessary analyses and transformations to parallel or sequentialapplications.

This chapter covers all the analyses implemented into the CPPC compiler, aswell as some of its most relevant internal implementation details.

3.1. Compiler overview

The CPPC compiler is built on the Cetus compiler infrastructure [37]. It iswritten in Java, which makes its code inherently portable. Although Cetus wasoriginally designed to support C codes, we have extended it to allow for parsing For-tran 77 codes as well. The specics of this extension are detailed in Section 3.4. Theimplementation uses the same basic intermediate representation language (IRL) forboth C and Fortran 77 codes, hence allowing the same analysis and transformationcode to be used with applications written in both languages.

63

64 Chapter 3. CPPC Compiler

Some of the analyses performed by the compiler require knowledge about proce-dure semantics (e.g. which procedure initializes the parallel system). We call thesetransformations semantic-directed. In order to provide the required semantic knowl-edge to the compiler, CPPC uses a catalog of procedures and its semantic behaviors.An example of the contents of this semantic catalog is shown in Figure 3.1, that ex-emplies information for the POSIX fopen and open functions. Both implementthe CPPC/IO/Open role (i.e. they open a le I/O stream). Both functions accept twoparameters: a path string and a le opening mode, and both return a le descrip-tor. Each procedure in the semantic catalog may implement any number of semanticroles. Each semantic role indicates that the function has a particular semantic be-havior. The implementation of a role by a function may have associated attributesindicating the specics of the semantic behavior. For instance, the DescriptorTypeattribute for the fopen/open case indicates the type of the returned descriptor. Itis internally used by the lesystem module in the portability layer of the CPPC li-brary to select a specic le stream manipulation routine, depending on the returneddescriptor type.

The catalog also contains data ow information. Procedure parameters are cate-gorized into input, output and input-output, depending on whether they are read,written, or read and written by the procedure. These tags allow the compiler toperform optimized data ow analyses for live variable detection (see Section 3.2.5).However, data ow information is not necessary for proper operation. If missing, thecompiler will take the conservative approach of considering all function parametersto be of input type, which implies that their values need to be recovered before theprocedure is called.

The semantic catalog is a portable and extensible solution for providing seman-tic information, and enables transformations to work with dierent programminginterfaces for a given subsystem. Supporting PVM communications, for instance,only requires the addition of the corresponding semantic information. Note that thesemantic catalog is not created or modied by CPPC users, but provided togetherwith the compiler distribution. In the current CPPC distribution, the catalog in-cludes information about Fortran 77 built-in functions; common UNIX functions;and the MPI standard, in both C and Fortran 77 versions. This is enough to sup-port most HPC MPI application in *NIX environments. Table 3.1 summarizes allexisting semantic roles and their attributes.

3.1 Compiler overview 65

<function name="fopen">

<input parameters="1,2"/>

<semantics>

<semantic role="CPPC/IO/Open">

<attribute name="FileDescriptor"

value="return"/>

<attribute name="Path" value="1"/>

<attribute name="DescriptorType"

value="CPPC_UNIX_FILE"/>

</semantic>

</semantics>

</function>

<function name="open">

<input parameters="1,2"/>

<semantics>

<semantic role="CPPC/IO/Open">

<attribute name="FileDescriptor"

value="return"/>

<attribute name="Path" value="1"/>

<attribute name="DescriptorType"

value="CPPC_UNIX_FD"/>

</semantic>

</semantics>

</function>

Figure 3.1: Semantic information for the fopen and open functions

66 Chapter 3. CPPC CompilerTable

3.1:Sem

anticroles

usedby

theCPPCcom

pilerRole

Semantic

Attrib

utes

CPPC/Comm/Initializer

Initializationof

theparallel

subsystemNone

CPPC/Nonportable

Non-p

ortablefunction

callNone

CPPC/IO/Open

Opening

ofale

stream

•FileDescriptor:How

isthe

descriptorreturned.•Path:Path

parameter.

•DescriptorType:

Typeof

thele

de-scriptor.

CPPC/IO/Close

Closing

ofale

stream•

FileDescriptor:

Descriptor

parame-

ter.

CPPC/Comm/Send

Message

send

•Blocking:true/false.

•Type:P2P

/Collective.

•Peer:Peer

rankparam

eter.•Tag:Tag

parameter.

•Communicator:Com

municator

param-

eter.•Request:Request

parameter,

fornon-

blockingcom

munications

only.CPPC/Comm/Recv

Message

receiveSam

eas

CPPC/Comm/Send.

CPPC/Comm/Wait

Com

munication

wait

•Blocking:true/false.

o•

Type:P2P

/Collective.

•Request:Request

parameter,

fornon-

blockingwaits

only.

CPPC/Comm/Ranker

Obtention

ofthe

rankof

aprocess

ina

communicator

•Rank:How

isthe

rankreturned.

CPPC/Comm/Sizer

Obtention

ofthe

sizeof

acom

municator

•Size:How

isthe

sizereturned.

3.2 Analyses and transformations 67

3.2. Analyses and transformations

The CPPC compiler is a multi-pass compiler. These kind of compilers processthe source code multiple times. Each pass takes the results of the previous oneas its input, and creates an intermediate output. The code is improved pass bypass, until the nal code is emitted. The rst of these passes is the parsing of thesource code, that builds the Cetus AST1. Once this is done, the compiler begins thesequential application of the set of passes that perform the core of the CPPC analysesand transformations for checkpointing instrumentation. When information must betransferred between passes, the compiler uses dierent forms of AST annotation.The following list of compiler passes is exhaustive, and described in the order thecompiler applies them:

3.2.1. CPPC initialization and nalization routines

The CPPC_Init_configuration() call initializes the CPPC conguration, whileCPPC_Init_state() marks the point for state initialization, creating all necessarymemory structures, such as lists of registered variables, execution contexts, etc.When resuming the execution of a parallel application, interprocess communica-tions are performed in order to nd the recovery line. Therefore, the state mustbe initialized after the communications system (e.g. after MPI_Init() for MPI ap-plications). The insertion of CPPC_Init_configuration() is straightforward, andperformed at the rst line of the application's entry procedure. The transforma-tion for inserting CPPC_Init_state() is semantic-directed. The compiler looks fora procedure implementing the CPPC/Comm/Initializer role. If found, it insertsthe state initialization right after it. Otherwise, the application is presumed to besequential or to use a communication system not needing initialization, and the callis inserted after CPPC_Init_configuration().

The nalization function, CPPC_Shutdown(), is inserted at exit points in theapplication code. It ensures that ongoing checkpoint operations are correctly nishedbefore the application ends.

1An Abstract Syntax Tree (AST) is a tree representation of the syntax of some source code.Each node of the tree denotes a construct occurring in the source code. The tree is abstract in thesense that it may not represent some constructs that appear in the original source (e.g. groupingparenthesis, which are not needed since the grouping of operands is implicit in the tree structure).

68 Chapter 3. CPPC Compiler

CPPC JUMP LABEL

CPPC_Create_call_image( "MPI_Comm_split", line_number );

CPPC_Register_parameter( &color, 1, CPPC_INT, "color", CPPC_STATIC );

CPPC_Register_parameter( &key, 1, CPPC_INT, "key", CPPC_STATIC );

CPPC_Commit_call_image();

MPI_Comm_split( comm, color, key,comm_out );

CONDITIONAL JUMP TO NEXT REC

Figure 3.2: Example of a non-portable procedure call transformation

3.2.2. Procedure calls with non-portable outcome

The CPPC library recovers non-portable state by means of the re-execution of thecode responsible for creating such state in the original run. The CPPC/Nonportablesemantic role is used for identifying procedures that create or modify non-portablestate. Upon discovery of a non-portable call, the compiler performs the transforma-tion depicted in Figure 3.2 for an MPI_Comm_split() call, where the CPPC instru-mentation added by the compiler is marked in bold. The block begins with a labelannotation which marks the jump destination from the previous REC, and ends witha conditional jump annotation, to mark the point where a jump to the next RECmust be performed. Both annotations are objects in the cppc.compiler.ast pack-age. The compiler also inserts a register for each input or input-output parameterpassed to the call that is a portable object. For more information on the runtimebehavior of non-portable calls and their instrumentation refer to Section 2.1.3.

3.2.3. Open les

In order to identify le opening calls, a semantic role CPPC/IO/Open is used. Thisrole accepts the following attributes:

FileDescriptor: A numeric value, or the special value return. It tells thecompiler which of the procedure parameters holds the le descriptor after thecall, or if it is available as the returned value.

Path: A numeric value indicating the parameter that contains the le path.

DescriptorType: A special constant value that must be dened and acceptedby the lesystem module in the portability layer. Current accepted values

3.2 Analyses and transformations 69

CPPC JUMP LABEL

fd = open( path, mode );

CPPC_Register_descriptor( desc_id, &fd, CPPC_UNIX_FD, path );

CONDITIONAL JUMP TO NEXT REC

Figure 3.3: Pseudocode for a le opening transformation

CPPC_Unregister_descriptor( &fd );

close( fd );

Figure 3.4: Pseudocode for a le closing transformation

are CPPC_UNIX_FD, indicating that this is an integer like the ones returnedby open, and CPPC_UNIX_FILE, that identies FILE * descriptors as the onesreturned by fopen.

The compiler uses this information to perform the transformation depicted inFigure 3.3, where instrumentation code is emphasized in bold. The desc_id pa-rameter passed to the descriptor registration call is a unique identier for this de-scriptor, automatically assigned by the compiler. When the application is recov-ered, CPPC_Register_descriptor() ensures correct repositioning of the le streampointer.

File closing calls are identied by the CPPC/IO/Close role, parameterized onlywith the FileDescriptor attribute dened as above. The transformation performedby the compiler is shown in Figure 3.4. Note that this block is not enclosed bycontrol ow annotations, because it does not need to be re-executed for proper staterecovery.

3.2.4. Conversion to CPPC statements

Some of the transformations to be performed from this point require statementsto be intensively annotated. However, the Cetus class for representing a statementdoes not include any means for this. Thus, a cppc.compiler.ast.CppcStatement

class was created, through the specialization of the original cetus.hir.Statement.It acts as a proxy for a contained regular statement, and adds some annotationstructures, such as data ow information (i.e. variables that are read or written bythe statement), information about whether a particular statement is a safe point,

70 Chapter 3. CPPC Compiler

etc. This transformation makes a pass over the entire application code substitut-ing regular Cetus statements for this CPPC version, enabling information sharingamongst subsequent transformations.

3.2.5. Data ow analysis

In order to identify the variables needed upon application restart, the compilerperforms a live variable analysis. This is a somehow complementary approach tomemory exclusion techniques used in sequential checkpointers to reduce the amountof memory stored, such as the one proposed in [48].

A variable x is said to be live at a given statement s in a program if there is acontrol ow path from s to a use of x that contains no denition of x prior to its use.The set LVin of live variables at a statement s can be calculated using the followingexpression:

LVin(s) = (LVout(s)−DEF (s)) ∪ USE(s)

where LVout(s) is the set of live variables after executing statement s, and USE(s)

and DEF (s) are the sets of variables used and dened by s, respectively. Thisanalysis, that takes into account interprocedural data ow, is performed backwards,being LVout(send) = ∅, and send the last statement of the code.

Before each checkpoint statement ci, the compiler inserts annotations to registerthe variables that must be stored in the checkpoint le, which are those containedin the set LVin(ci). The data type for the register is automatically determined bychecking the variable denition. Variables registered or dened at previous check-points are not registered again. Also, before each checkpoint ci, the compiler insertsunregister annotations for the variables in the set LVin(ci−1)−LVin(ci), the set ofvariables that are no longer relevant.

Checkpoints can be placed inside any given procedure. For a checkpoint state-ment ci, let us dene:

Bci= s1 < s2 < . . . < send

as the ordered set of statements contained in all control ow paths from ci (excludingci) and up to the last statement of the program code, where the < operator indicatesthe precedence relationship between statements. Note that, if a checkpoint is placedinside a procedure f , not all statements in the set Bci

will be inside f . Let us

3.2 Analyses and transformations 71

denote by Bfci

= s1 < . . . < sn the ordered set of statements contained inside f ,and Lf

ci= Bci

−Bfcithe ordered set of statements left to be analyzed outside f .

The interprocedural analysis and register insertion is performed according to thefollowing algorithm:

For a checkpoint statement ci contained in a procedure f , the live variableanalysis is performed for the set Bf

ci, and registers for locally live variables are

inserted before ci.

For the set Lfci, containing the statements that are left unanalyzed in the

previous step, let us consider g to be the procedure containing the statementsn+1. The statement executed immediately before sn+1 must be a call to f .Note that Lf

ci= Bg

cit Lg

ci.

The live variable analysis is performed for the set Bgci, and registers for locally

live variables are inserted before the call to f .

The process is repeated for the statements contained in the ordered set Lgci.

This algorithm ensures that, upon application restart, all variables will be denedbefore being used, and thus the portable state of the application will be correctlyrecovered.

Proof of Correctness: Let us consider a variable v which appears in the state-ments contained in Bci

, and let sv ∈ Bcibe the statement where it rst appears.

There are three dierent cases to analyze: v can either be an input, an output,or an input-output variable for sv. Using the denition of the live variable analysis:

input: v ∈ USE(sv) ∧ v /∈ DEF (sv)

USE(sv) ⊂ LVin(sv)⇒ v ∈ LVin(sv)

Since sv was the rst appearance of v in Bci, there is no previous statement

which denes its value, meaning that v belongs to the live variable set forevery statement before sv:

@i/(v ∈ DEF (si)) ∧ (si < sv)⇒

v ∈ LVin(sj),∀j/sj < sv

72 Chapter 3. CPPC Compiler

In particular, since ci < sv: v ∈ LVin(ci).

A register will be inserted before sv when analyzing its containing procedure.This register will generate v's value when restarting the application.

output: v /∈ USE(sv) ∧ v ∈ DEF (sv)

In this case, it is guaranteed that v /∈ LVin(sv).

Since sv was the rst appearance of v in Bci, there is no previous statement

which uses its value, meaning that v does not belong to the live variable setfor every statement before sv:

@i/(v ∈ USE(si)) ∧ (si < sv)⇒

v /∈ LVin(sj),∀j/sj < sv

In particular, since ci < sv: v /∈ LVin(ci).

No register will be inserted; v's value will be generated upon reaching sv,therefore dening it.

input-output: v ∈ USE(sv) ∧ v ∈ DEF (sv)

This case is similar to the input one.

Proof of Termination: The algorithm terminates if all statements containedinto Bci

are analyzed. Bciis a nite set, since its maximum number of elements

is equal to the total number of statements in the application being analyzed. Theevolution of the cardinality of the unanalyzed set of statements is as follows:

In the rst phase of the algorithm (the analysis of procedure f), the statementsin Bf

ciare analyzed. Either s1 ∈ f and #Lf

ci< #Bci

, or ci is the last statementin f and #Lf

ci= #Bci

.

In each of the subsequent phases, the set of unanalyzed statements can bewritten as:

Lpnci

= spn

1 , . . . , spnm = Bci

− (n⊔

j=1

Bpjci

)

3.2 Analyses and transformations 73

where each pj is the procedure being analyzed in phase j. In phase n, thealgorithm analyzes the procedure containing the statement spn

1 . Therefore, atleast one statement is analyzed, and #Lpn+1

ci< #Lpn

ci.

Since all the statements must be contained in a procedure in order to be in theinitial Bci

and this is a nite set, the termination of the algorithm in a nite numberof steps is guaranteed.

The compiler does not currently perform optimal bounds checks for pointer andarray variables. This means that some arrays and pointers are registered in a conser-vative way: they are entirely stored if they are used at any point in the re-executedcode.

When dealing with calls to precompiled procedures located in external libraries,the default behavior is to assume all parameters to be of input type. Therefore,registration calls will be inserted for the previously unrecovered variables containedin the set LVin(sp), being sp the analyzed procedure call. To avoid this defaultbehavior, the CPPC compiler can use data ow information available in the semanticcatalog, as explained in Section 3.1.

3.2.6. Communication analysis

In order to automatically nd regions in the code were neither in-transit norinconsistent communications exist, that is, safe points, the compiler must analyzecommunication statements and match sends to their respective receives. The ap-proach the compiler uses is similar to a static simulation of the execution. Twocommunications are considered to match if two conditions hold:

1. Their tags are the same or the receive uses MPI_ANY_TAG.

2. Their sets of sources/destinations are the same: if a process i sends a messageto a process j, then j must execute the corresponding receive statement usingi as source.

In order to statically compare tags and source/destination pairs, the compilerperforms constant propagation and folding to determine their literal values during

74 Chapter 3. CPPC Compiler

the execution. Since propagating and folding all application constants may prove tobe as computationally intensive as the actual execution, the compiler restricts thisoperation to the set of variables that are involved in the calculations of parametersaecting communications. The set of communication-relevant variables is recursivelycalculated as:

1. Variables directly used in tags or source/destination parameters.

2. Variables on which a communication-relevant variable depends.

In order to optimize this process, only statements that modify communication-relevant variables are analyzed for constant folding, since any other does not aectcommunications. The data ow information is obtained from the annotations intro-duced by the previous compilation pass. Some communication-relevant variables aremultivalued, that is, they potentially have a dierent value for each process. Thus,the compiler needs to know how many processes are involved in the execution of thecode to perform the constant folding. For instance, many applications are writtenin a scalable fashion, using code that dynamically calculates neighbor processes inthe commmunication topology. These calculations are aected by the total numberof processes involved in the parallel execution. The number of processes is providedto the compiler via command-line parameters. Then, using the CPPC/Comm/Rankerand CPPC/Comm/Sizer semantic roles, the compiler is able to nd values for thevariables that contain the number of processes and the rank of a process. Note thatthe latter is multivalued (e.g. 0, . . . , N − 1 if there are N processes in the exe-cution). Expressions derived from multivalues are, in the general case, multivaluedthemselves. The constant folding and propagation is performed together with thecommunication matching in the same compilation pass. The compiler is able tofold array expressions and correctly calculate array values statically, as long as theydepend entirely on values known at compile time.

For keeping track of the communications status, the compiler uses a buer ob-ject, which starts out empty. The analysis begins at the application's entry point.Statements that are neither control ow- nor communication-related are ignored.Each time the compiler nds a new communication, it rst tries to match it withexisting ones in the buer. If a match is not found, the communication is addedto the buer and the analysis continues. If a match is found, both statements areconsidered linked and removed from the buer, except when matching non-blockingsends and receives, in which case they remain in the buer in an unwaited statusuntil a matching wait is found.

3.2 Analyses and transformations 75

A statement in the application code will be considered a safe point if, and only if,the buer is completely empty when the analysis reaches that statement. An emptybuer implies that no pending communications have been issued, and therefore it isimpossible for an in-transit or inconsistent message to exist at that point.

The following subsections deal with particular aspects of the communicationsanalysis, such as analyzing conditional expressions, procedure calls, collective com-munications and the limitations of this approach.

• Conditional statements

Conditional statements need special treatment, since their execution order isnondeterministic. Upon reaching a conditional, SPMD processes potentially divideinto two groups: those executing the true clause, and those executing the false one.The compiler uses linked buers to analyze conditional statements. The dierencewith a regular buer is that a linked buer Bl is associated to another one Bo in away that, for a statement to be considered a safe point, not only the Bl has to beempty, but Bo must also be consistent with the safe point condition. Note that thismay be generalized as a chain of linked buers which check if they are themselvesempty and delegate the nal decision on their link until one is not empty or the rootone is reached, much like in a chain of responsibility pattern [28].

The compiler creates an empty buer Bt linked to the old one Bo, and uses it toanalyze the true clause. Found communications are added to Bt, and matched asusual, with an important dierence: each one is marked as being executed only ifthe expression controlling the conditional statement execution holds. After the trueclause is analyzed, the compiler proceeds in the same way with the false one using adierent buer Bf also linked to Bo. This time, each communication is marked asbeing executed only if the expression controlling the conditional does not hold.

When both branches have been fully analyzed, the compiler has three dierentbuers, Bt, Bf and Bo, that need to be merged into a single one. First, redundantcommunications are removed from Bt and Bf . Although both conditional paths ex-ecute a dierent set of statements, equivalent communications may exist. These arecommunications that are executed in dierent statements in the code but matchedto a single one outside the conditional. If these redundant communications werenot identied, the algorithm would yield incorrect results, since only one of theequivalent communications would be correctly matched. Thus, two communication

76 Chapter 3. CPPC Compiler

statements are considered equivalent if they both match the same set of statementsaccording to their source/destination pair and tag, and if both are issued underincompatible control expressions (e.g. on dierent branches of the same conditionalstatement). Redundant communications are substituted by a single representativeone, that will be matched at the same point in the code where each of the redundantones would during execution. This enables consistent discovery of safe points in thepresence of redundant communication statements.

The next step consists in merging all three buers into a single one that rep-resents the communications state after executing the conditional. The process ofmerging two buers is a binary operation. The left hand side buer (Blhs) repre-sents the communications state before executing a block of code. The right hand sidebuer (Brhs) represents the communications issued by the execution of that blockof code. The communications in Brhs are orderly injected into Blhs. Non matchingcommunications are added to Blhs. Matches are dealt with as previously explained.When the process ends, the resulting buer represents the communications state ofa process that, starting in a state represented by Blhs, executes the block of codeissuing communications in Brhs. Note that this process implies a deferral of thematches for the statements in the new buer. Since there is no correct partial orderfor the statements in conditional branches, both are independently analyzed andthen mixed before being injected into the buer representing the old state.

First, Bt and Bf are merged together. The result is merged with Bo. The resultof this second merge is used for the analysis of subsequent communications. Thereason for merging the buers obtained from analyzing the conditional rst is thatoften communications in dierent conditional branches match between themselves.

Note that some conditional control expressions may be calculated during theconstant folding analysis. In these cases, the compiler analyzes only the appropriateconditional branch. The compiler is also able to detect situations in which thecontrol expression is multivalued, and deal with them accordingly.

• Procedure calls

When the compiler nds a procedure call, it stops the ongoing analysis and movesto analyze the code of the called procedure, using the same communications buer.However, communications issued inside the procedure are also cached separatelyfor optimization purposes. If the procedure does not modify any communication-

3.2 Analyses and transformations 77

relevant variable, cached results can be reused when a new call to the same procedureis found, without analyzing the procedure code, but just symbolically analyzing theirtags and source/destination parameters again.

The compiler employs this optimization whenever possible, thus avoiding therepeated analysis of procedures each time a call to them is found. If the proceduremodies any communication-relevant variable, then its code is analyzed each time acall is found, since these modications have to be tracked and applied to guaranteecorrect constant folding.

• Collective communications

In SPMD applications, the usual way of coding collective communications is tomake all involved processes execute a single, non conditional line of code. Beforeand after the execution of that line the collective communication does not aect theconsistency of communications. However, in the general case, a code may containa collective communication spread across several conditional paths. In this case,all its parts must be detected as redundant and the compiler must ensure thatno checkpoints are inserted such that some processes take its local one before thecollective while some others do so after it.

• Limitations

Some parallel applications present nondeterministic or irregular communicationpatterns (those depending on runtime input data, where tags and/or source/desti-nation parameters are derived from the input data and cannot be statically deter-mined). In these situations, the compiler must apply either a conservative approachor heuristics in order to detect safe points. A conservative approach involves match-ing communications to their latest possible match in the code, although this can leadto the compiler not nding safe points at points selected by the checkpoint insertionanalysis described in the next section. The heuristic currently in use is to considerthat two communications match if no matching incompatibility between their tagsand source/destination pairs is found.

78 Chapter 3. CPPC Compiler

3.2.7. Checkpoint insertion

The compiler locates sections of the code that take a long time to execute, wherecheckpoints are needed to guarantee execution progress in the presence of failures.Since computation time cannot be accurately predicted at compile time withoutknowledge of the computing platforms and input data, heuristics are used. Thecompiler discards any code location that is not inside a loop, and ranks all loopnests in the code using computational metrics. Currently, a metric derived fromboth the number of statements executed inside the loop and the number of variablesaccessed in them is used. A loop might have many statements but accessing mostlyconstants. These are usually not good candidates for checkpointing, since theyinvolve initializing variables and their execution does not last long. In this case,considering the number of variable accesses in the loop reduces the cost associatedto such loops, thus making them less prone to be checkpointed.

Let l be a loop in L, the loop population of the application P , and s and a twofunctions that give the number of statements and variable accesses in a given blockof code, respectively. Let us dene S(l) = s(l)/s(P ) and A(l) = a(l)/a(P ) the totalproportion of statements and accesses, in that order, that exist inside a given loopl. The heuristic complexity value associated to each loop l is calculated as:

h(l) = −log(S(l) · A(l)) (3.1)

Eq. (3.1) multiplies S(l) and A(l) to ensure that the product is bigger for loops thatare signicant for both metrics. It applies a logarithm to make variations smoother.Finally, it takes the negative of the value to make h(l) strictly positive. Thus, alower value implies that more computing time is estimated for that loop. In orderto select the best candidates for checkpoint insertion, the compiler ranks loops inL attending to h(l) and applies thresholding methods [57]. After exploring severalpossibilities, a two-step method has been adopted. First, a shape-based approachis used to select the subset of the time-consuming loop nests in the application.The second step improves this selection by means of a cluster-based technique.Figures 3.5 and 3.6 show the proposed method applied to the BT application of theNAS Parallel Benchmarks [3].

In the rst step, using the so-called triangle method, loops in L are divided intotwo classes: time-consuming loops and negligible loops. Conceptually, the trianglemethod consists in: (a) drawing a line between the rst and last values of h(l);(b) calculating the perpendicular distance d(l) to this line for all l ∈ L; and (c)

3.2 Analyses and transformations 79

Figure 3.5: First step: shape-based threshold

calculating a threshold value lt such that d(lt) = maxd(l). This rst step selects asubset of the loop population, referred to as H, as time-consuming loops candidatesfor checkpoint insertion, as shown in Figure 3.5. The triangle method was originallydeveloped in the context of image processing [73], and appears to be especiallyeective when there is a narrow peak in the histogram, which is often the case forthe proposed h(l) function in real applications.

However, this rst threshold selects more loops than would be desirable for check-point insertion. The second step of the proposed algorithm renes this selection us-ing a cluster-based thresholding algorithm. Loops in H with similar associated costsare grouped into clusters, Ci, built by selecting the local maximums of the secondderivative of h(l) as partitioning limits. Since h(l) is monotonically increasing, localmaximums in h′′(l) represent inection points at which h(l) begins to change moresmoothly. For an application with k clusters, the method calculates a thresholdvalue t such that:

t∑i=0

h(l0i+1)− h(l0i ) >k−1∑

i=t+1

h(l0i+1)− h(l0i ) (3.2)

where l0i denotes the rst loop in Ci. This method selects for checkpoint insertion allthe loops inside the clusters Ci such that i ≤ t. In the example shown in Figure 3.6,loops in subset H are grouped into two clusters, delimited by the single maximum

80 Chapter 3. CPPC Compiler

Figure 3.6: Second step: cluster-based threshold

in h′′(l). Applying Eq. (3.2), the compiler selects cluster C0 for checkpoint insertion,which contains only one loop (which corresponds to the main computational loop inBT).

Once the loops in which checkpoints are to be inserted are identied, the com-piler uses the communication analysis described in the last subsection to insert acheckpoint at the rst available safe point in each selected loop nest. Experimentalresults are detailed in Chapter 4.

• Manual checkpoint insertion

Instead of, or in addition to, automatically placed checkpoints, the CPPC com-piler allows for manual insertion of checkpoints through two dierent user-inserteddirectives:

#pragma cppc checkpoint: This directive orders the compiler to place acheckpoint call in the exact position where it is inserted.

#pragma cppc checkpoint loop: This directive is placed before a loop, andtells the compiler to insert a checkpoint at the rst available safe point insideit.

3.2 Analyses and transformations 81

DO I=A,B,STEP

!$CPPC CHECKPOINT

...

END DO

CPPC_INIT_IT = A

CPPC JUMP LABEL

CPPC_REGISTER( CPPC_INIT_IT, 1,

CPPC_INT, "CPPC_INIT_IT" );

DO I=CPPC_INIT_IT,B,STEP

CPPC_INIT_IT=I

!$CPPC CHECKPOINT

...

END DO

Figure 3.7: Modications to a checkpointed Fortran 77 loop

Cetus represents compilation directives through its cetus.hir.Annotation class,that encapsulates a String object that contains the annotation text. Instead,the CPPC compiler denes a cppc.compiler.ast Java package, dening severalAnnotation subclasses that are internally used for inter-pass communications. Thisachieves strong typing of the CPPC annotations, allowing for the compiler to uni-formly nd and process them using object-oriented techniques such as method over-loading. An initial pass translates user-inserted checkpoint directives to the internalCPPC representation.

The Fortran 77 syntax for inserting manual directives is !$CPPC CHECKPOINT

and !$CPPC CHECKPOINT LOOP.

3.2.8. Language-specic transformations

Some language-dependent modications are necessary in order for the restartprotocol to consistently work. Particularly, restarting a Fortran 77 execution froma checkpoint inside a loop needs the loop bounds to be changed in order for theexecution to be resumed from the correct loop iteration. Fortran 77 does not acceptloop indexes to be dynamically changed once the control ow has entered the loop.Instead, the set of values for the iteration variable is constructed when enteringthe loop, and its value is updated at each iteration regardless of internal changes.Therefore, a modication to the loop variable persists only until the next iterationstarts, where it is modied again and the modication lost.

Figure 3.7 depicts the changes made to a Fortran 77 loop in order to enableinternal checkpointing. During a regular execution, the start value A is assigned to a

82 Chapter 3. CPPC Compiler

CPPC-introduced variable called CPPC_INIT_IT, which is updated at each iterationof the loop before the checkpoint call. Thus, it is guaranteed that it will be storedin the state le containing the value of the checkpointed iteration. When the ap-plication is restarted, this value will be recovered through the variable registration,and the original loop bounds altered, starting the iteration at the correct value.

3.2.9. Code generation

At this point, all required analyses have been performed, and the code has beenannotated with objects in the cppc.compiler.ast package. This last pass looksfor those annotations and replaces them with their corresponding code in the AST.Then, a code generator reads the AST and generates the output in the originalprogramming language.

3.3. Case study

For illustrative purposes, an example of how CPPC inserts checkpointing codeinto a parallel application is presented. The target application is a simplied Cversion of a large-scale engineering code that performs a crack growth simulationusing the Dual Boundary Element Method (DBEM) [30], further used in Chapter 4for the experimental evaluation of the CPPC tool. Figure 3.8 shows the originalcode, which can be divided in ve basic sections:

1. Reading of the execution parameters from a conguration le (done by theprocess with rank 0).

2. Creation of the necessary MPI structures, such as communicators and datatypes for supporting the communication between processes.

3. Distribution of the execution parameters and data previously read by process0 among all the application processes.

4. Call to the frmtrx() subroutine, responsible for building the system matrix(variable a) and the right-hand side vector (variable b) of the equation systemused for modelling the engineering problem.

3.3 Case study 83

int main( int argsc, char ** argsv ) // Variable definitionsMPI_Init( &argsc, &argsv );MPI_Comm_rank( MPI_COMM_WORLD, &myrank );

// Host reads parametersif( myrank == 0 )

iinput = fopen( "control.dat", "r" );fread( iinput, "%s\n", outname );fread( iinput, "%d\n", &numstep );...

// Creation of communication topologyMPI_Cart_create( MPI_COMM_WORLD, ndims,

dims, periods, reorder, &cart_comm );...

// Distribution of parameter dataMPI_Bcast( &nel, 1, MPI_INTEGER, 0,

cart_comm );MPI_Bcast( &nnp, 1, MPI_INTEGER, 0,

cart_comm );...

// Construction of system matrix and rhsfrmtrx();

// Solution of equation systemsolve();

...

void solve() // Variable definitions...for( i=0; i < maxiter; i++ )

mpi_dgemv( ml, nl, a, xm, ax, cart_comm );mpi_dcopy( ml, b, 1, r, 1, cart_comm );mpi_daxpy( ml, -1, ax, 1, r, 1, cart_comm );...

...

Figure 3.8: Case study: original DBEM code

84 Chapter 3. CPPC Compiler

int main( int argsc, char ** argsv ) // Variable definitionsvoid * cppc_jump_points[4] =

&&CPPC_EXEC_1, &&CPPC_EXEC_2, ...int cppc_next_jump_point = 0;MPI_Init( &argsc, &argsv );CPPC_Init( &argsc, &argsv );// Conditional jump to CPPC_EXEC_0if( CPPC_Jump_next() )

goto * cppc_jump_points[ jump_index++ ];MPI_Comm_rank( MPI_COMM_WORLD, &myrank );

// Host reads parametersCPPC_EXEC_0:

if( myrank == 0 ) // Conditional jump to CPPC_EXEC_1 (code omitted)iinput = fopen( "control.dat", "r" );

CPPC_EXEC_1:CPPC_Register_descriptor( 0, iinput,

CPPC_UNIX_FILE, "control.dat" );// Conditional jump to CPPC_EXEC_2 (code omitted)fread( iinput, "%s\n", outname );...

// Creation of communication topologyCPPC_EXEC_2:

CPPC_Register_parameter( &ndims, ..., CPPC_STATIC );dims = CPPC_Register_parameter( dims, ..., CPPC_DYNAMIC );periods = CPPC_Register_parameter( periods, ..., CPPC_DYNAMIC );CPPC_Register_parameter( &reorder, ..., CPPC_STATIC );MPI_Cart_create( MPI_COMM_WORLD, ndims,

dims, periods, reorder, &cart_comm );...// Conditional jump to CPPC_EXEC_3 (code omitted)

// Distribution of parameter dataMPI_Bcast( &nel, 1, MPI_INTEGER, 0, cart_comm );MPI_Bcast( &nnp, 1, MPI_INTEGER, 0, cart_comm );...

// Construction of system matrix and rhsfrmtrx();

// Solution of equation systemCPPC_EXEC_3:

CPPC_Context_push( "solve", 0 );solve();CPPC_Context_pop();...CPPC_Shutdown();

Figure 3.9: Case study: CPPC-instrumented DBEM code

3.3 Case study 85

void solve() // Variable definitions + CPPC definitions// Conditional jump to CPPC_EXEC_4 (code omitted)...

CPPC_EXEC_4:CPPC_Register( &i, ..., CPPC_STATIC );CPPC_Register( &ml, ..., CPPC_STATIC );a = CPPC_Register( a, ..., CPPC_DYNAMIC );xm = CPPC_Register( xm, ..., CPPC_DYNAMIC );ax = CPPC_Register( ax, ..., CPPC_DYNAMIC );b = CPPC_Register( b, ..., CPPC_DYNAMIC );...// Conditional jump to CPPC_EXEC_5 (code omitted)for( i=0; i < maxiter; i++ )

CPPC_EXEC_5:CPPC_Do_checkpoint( 0 );mpi_dgemv( ml, nl, a, xm, ax, cart_comm );mpi_dcopy( ml, b, 1, r, 1, cart_comm );mpi_daxpy( ml, -1, ax, 1, r, 1, cart_comm );...

...

Figure 3.10: Case study: CPPC-instrumented DBEM code (cont.)

5. Call to the solve() subroutine, containing the computational loop whichsolves the system.

The resulting code after processing by the CPPC compiler can be seen in Fig-ure 3.9 and Figure 3.10. The compiler inserts new variables which are used to controlthe program ow on application restart, ensuring that only relevant blocks of codeare executed. The applied transformations are the following:

1. The compiler detects that the MPI_Init() call implements the CPPC/Comm/-

Initializer role. Therefore, the CPPC_Init() call is placed right after it.The nalization function is inserted at the end of the code.

2. The compiler identies the fopen() function as implementor of the CPPC/-

IO/Open role and inserts the C version of the le handling function, CPPC_-Register_descriptor().

3. The compiler nds that the MPI_Cart_create() subroutine is tagged withthe CPPC/Nonportable role. It inserts the appropriate CPPC_Register_-

parameter() calls and adds necessary ow control structures.

86 Chapter 3. CPPC Compiler

4. The compiler performs a safe point detection and inserts a checkpoint in theloop in the solve() subroutine. It also performs a live variable analysis andintroduces the relevant registers prior to its execution. These include the i, ml,a, xm, ax and b variables. nl is not registered, since it is declared as a constantinitialized in the code. r is an output parameter for the mpi_dcopy() call, andhence it is not live at the analysis point. The insertion of a checkpoint insidesolve() causes calls to this subroutine to be re-executed on restart. Contextpush and pop function calls are placed enclosing such calls to notify the CPPClibrary of context changes.

If the application is restarted after a failure, the program ows through relevantblocks recovering the entire application state. Variables are initialized and both theMPI and CPPC initialization functions are called. Then, a conditional jump tolabel CPPC_EXEC_0 takes place, transferring execution to the conditional structurethat performs the opening of the le 'control.dat'. Note that the condition istested before proceeding to REC execution, and so the semantics of the code arepreserved. The CPPC_Register_descriptor() function ensures that the iinput

descriptor state is correctly recovered. Next, the communication topology is recre-ated upon reaching label CPPC_EXEC_2. Then the execution moves to the subrou-tine call on label CPPC_EXEC_3. Once inside, relevant variable values are recoveredupon reaching label CPPC_EXEC_4, and later the checkpoint on label CPPC_EXEC_5 isreached. CPPC determines that all the state has been restored, and the executionis resumed normally in checkpoint operation mode.

3.4. Implementation details

This section covers the compiler implementation in some detail. Each subsectionis focused on a dierent implementation issue, covering the addition of the necessaryfeatures for Cetus to support Fortran 77 codes; the design patterns used to share asmuch code as possible between the C and Fortran 77 versions of the compiler; andthe elegant implementation of code analyses.

3.4.1. Fortran 77 support

The original Cetus provides a compiler research environment that facilitates thedevelopment of interprocedural analyses for C applications. There are three basic

3.4 Implementation details 87

layers on the Cetus design: a C parser, that reads source code and creates itsassociated AST; classes representing the AST itself, providing representations of thecode and operations to manipulate it; and a code generator back-end that performsthe reverse transformation, reading an AST and outputting a C code.

Our goal when implementing the CPPC compiler was to be able to reuse thecompilation passes' code for processing both C and Fortran 77 applications. Inorder to do so, a front-end capable of building a Cetus-compatible AST repre-sentation of a Fortran 77 code was required. This front-end was implemented bywriting a Fortran 77 grammar using the ANTLR [46] parser generator tool. Itwas also necessary to add new classes to the Cetus AST to represent Fortran 77concepts that do not exist in C. The following abstractions were included on thecppc.compiler.ast.fortran77 package:

COMMON declarations: A declaration with a block name, and an associated listof regular declarations included in the common block.

DATA declarations: A declaration containing a list of variables and their asso-ciated initializers.

DIMENSION declarations: A declaration containing a list of variables and theirassociated dimensions.

IMPLICIT declarations: A class containing a beginning character, an end char-acter, and an associated specier to the range.

COMPLEX datatype: A specier for representing complex datatypes.

Array speciers: Fortran array speciers dier from C ones in that their lowerbound may be specied when the array is declared. Since the Cetus versionof the array specier only contains an upper bound, this class was created toallow for the inclusion of lower bound declarations.

String speciers: Fortran allows for strings to be dened as CHARACTER* vari-ables. Since the concept of string does not exist in C, this class was added.

Double precision literals: Cetus does only support oating point literals, with-out specifying their precision. Therefore, all literals should be converted eitherto oats or doubles when building the AST, which would lead to potentialprecision issues. This class is used for representing double literals, while theoriginal Cetus one is reserved for oats.

88 Chapter 3. CPPC Compiler

Substring expressions: Supports substring expressions such as STRING(A:B).

Implied DO: Supports implied do constructions such as print *,(i+1,i=0,N).

Intrinsic I/O calls: Some Fortran intrinsics, such as print or write, accepta number of regular parameters and then any number of variable arguments,such as in write(FILE), i, j, k, ... This class represents such a call,separating its parameters in regular ones (to be included between brackets)and variable ones.

FORMAT statements.

Computed GOTOs: Support for Fortran-style computed GOTOs, of the form GOTO

(LABEL1,LABEL2,...) INDEX.

DO loops: Support for Fortran-style DO loops, that have no direct equivalent inC.

Ideally, once the source code is represented as an AST, the compilation passesare able to handle applications regardless of the source language they were originallywritten in. However, there are certain code structures that have to be handled indierent ways for dierent programming languages. The next subsection covers howthese dierences are handled.

Once all passes have been applied, a Fortran 77 back-end, created specicallyfor CPPC, reads the modied AST and generates the Fortran 77 output.

3.4.2. Sharing code between the C and Fortran 77 compilers

Since both the C and the Fortran 77 compiler use the same basic AST, mostof the code used for analyses and transformations is usable by the compilers ofboth programming languages. Some constructs, however, need a slightly dierenttreatment. In order to share as much code as possible, the implementation of thecompilation passes uses the template method design pattern [28]. Analyses thatneed dierent versions have a common implementation depending on some abstractmethods, which are then implented in language-dependent classes. The operationsthat dier and therefore need these pattern to be used are:

Insertion of control ow code: Two actions dier when inserting control owcode. First, the number and types of variables inserted to orchestrate the

3.4 Implementation details 89

restart jumps. While the C version uses an array of void * values, Fortran77-style computed gotos do not require such variable. Also, since the Fortran77 version of CPPC_JUMP_NEXT() is a subroutine rather than a function, itadds a CPPC_JUMP_TEST variable to store its results. Besides, the conditionaljumps code is dierent due to the dierences in computed gotos syntax.

Language-dependent transforms: As detailed in Section 3.2.8.

Pointers: All the transformations that depend on whether the accessed vari-ables are pointers or statically allocated are only present in the C version ofthe compiler. These kind of operations include determining whether a variableis static or dynamic, obtaining the size of a variable, and obtaining the basememory address for a variable.

CPPC calls: Some CPPC library functions have dierent signatures for Cand Fortran 77. Thus, the code that inserts these calls into the application isdierent for both compilers.

Analysis of language-specic AST objects: Some of the representations of thecode are specic to one of the languages. Examples of these are implied do

loops in Fortran 77, or for loops in C. The classes in charge of AST analysis,further described in the next subsection, are specialized so that superclassescontain analysis methods for common structures, while the subclasses containanalysis code for language-specic constructs only.

3.4.3. AST analyzers

Performing interprocedural analyses on the entire application is a complex pro-cess, that includes many interactions between the dierent levels of the abstractrepresentation. The AST is composed primarily of objects belonging to two classhierarchies: statements and expressions. Classes in charge of performing interpro-cedural analyses, such as those for performing the data ow analysis, traverse thetree in a depth-rst fashion, beginning with statements and eventually accessingtheir contained expressions. Thus, they need to have a method for processing eachdierent class in each of the hierarchies. The normal approach for solving thisdesign problem would be to use a visitor pattern [28]. However, this would re-quire both the statements and expressions hierarchies to be modied in order toaccept their corresponding visitors. To avoid modifying the Cetus core, a variation

90 Chapter 3. CPPC Compiler

Figure 3.11: AST analyzer implementation by instantiation of the method dispatcher

of the visitor pattern taking advantage of the Java reection interface was imple-mented. An example for a class analyzing statements is depicted in Figure 3.11.The MethodDispatcher is a an abstract Java Generic implementing a single oper-ation: accessing the Java reection interface to fetch a method able to process anobject of a given class t in the T hierarchy. AST analyzers are implemented throughinstantiation of this Generic. For instance, the StatementAnalyzer class binds thetemplate parameter T to the Cetus Statement superclass, and implements methodsfor analyzing each of its subclasses. The only provided public method accepts aCppcStatement, introduced in Section 3.2.4. It fetches the contained statement,and issues a dispatch() petition asking for the appropriate method for analyzingit. The dispatcher answers with the only protected method implemented by theStatementAnalyzer that accepts an object of the specied class of statement (e.g.DeclarationStatement, ExpressionStatement, etc.). Thus, visitors for each ofthe AST hierarchies are implemented without modications to the original codes.

However, accessing the Java reection interface is a very inecient operation.The compiler, in a typical analysis of a parallel code, may perform millions ofcalls to the dispatch function. In order to alleviate the reection overhead, theMethodDispatcher class includes an associative cache linking subclasses of T andtheir corresponding processing methods. Since the target hierarchies do not have alarge number of members the cache will quickly be lled, eectively neglecting the

3.5 Related work 91

penalization for using the reection interface.

3.5. Related work

Most checkpointing approaches in the literature perform structural analyses andsource code modications to instrument checkpointing insertion. Many of them,however, leave the insertion of checkpoints to be manually performed by the user,while automatically performing the remaining instrumentation (variable storage, re-covery, control ow, etc.). Porch [51] and C3 [1315] require that the user insertscheckpoint calls in the code. These calls will only trigger an actual checkpointaccording to a frequency timer. These potential checkpoints were originally in-troduced by CATCH GCC [39], which also automated their insertion. A potentialcheckpoint was introduced at the beginning of subroutines, and at the rst line insidea loop. This checkpoint placement guaranteed, in the general case, that potentialcheckpoint calls would be executed often enough so as to provide a checkpointingfrequency reasonably close to the desired one. However, this approach cannot be fol-lowed by CPPC, where checkpointing frequencies are not dened in temporal terms,due to the need to statically coordinate all processes independently of how long theytake to progress through the application's execution.

For checkpointing schemes that do not use runtime coordination, such as CPPC,checkpoints have to be inserted at places where the global consistency of the ap-proach is guaranteed. When working with implicitly parallel languages, the compileris able to use the native constructions for parallelism to extract information aboutsafe places for checkpointing. Such is the case of the work in ZPL by Choi et al. [20],where safe communication-free checkpoint ranges are detected, and a checkpoint in-serted at the single location in the range with the fewest live variables.

When working with explicit communication interfaces, such as MPI, the commu-nication pattern is user-dened. The compiler has to interpret the communicationstatements inserted in the code, and perform a communication matching to nd safelocations for checkpointing. Some analysis models have been developed with the goalof testing the correctness of the communication pattern in an application. However,pursuing correctness verication is a tougher problem than nding safe regions in thecode, and these models typically present strong limitations, such as not being ableto deal with non-blocking communications or wildcard receives (MPI_ANY_SOURCEand MPI_ANY_TAG) [59].

92 Chapter 3. CPPC Compiler

Performing static communication analyses typically involves the use of MPI con-trol ow graphs [58], that extend the concept of a control ow graph with theintroduction of communication edges that connect communication nodes. Since cor-rectness tests are not one of their goals, they assume that the communications arecorrect and conservatively represent indeterminacies by drawing all possible commu-nication edges between the nodes involved. A node in the graph is considered to be asafe point if there is no communication edge connecting one of its ancestors with oneof its successors. CPPC uses the same approach as described for the analysis, exceptthat no graph is used. Since our goal is not to obtain a representation of the com-munication pattern in the application, but rather, to simply categorize points in thecode as either safe or non-safe, the graph may be ommitted, and the communicationbuer object previously described used for representing pending communications ateach point in the code. Not building the graph has several advantages in termsof eciency. Mainly, it minimizes memory consumption and does not require theanalysis of the graph resulting from the compilation pass.

With regards to the automatic checkpoint insertion, some theoretical approachescalculating the optimal mathematical solution to the checkpoint placement problemexist [67, 69]. However, these assume that all involved parameters are known. Thisis not the case under CPPC, where the execution environment is not assumed to bexed due to its inherent characteristic of portability.

Checkpoint le sizes are tightly related to scalability issues when checkpointingparallel applications. Thus, dierent approaches to reduce the amount of data storedin state les have been studied during the last decade. Plank et al. proposed in [48] amemory exclusion technique based on performing a dead variable analysis to identifymemory regions which can be safely excluded from the checkpoint process. CPPCimplements a similar technique, based on performing a live variable analysis atcompile time to identify those variables that should be included in the checkpointle.

3.6. Summary

This chapter has covered both the design and implementation of the CPPCsource-to-source compiler, in charge of automating the integration of a parallel orsequential application with the CPPC library. It is implented in Java, oeringwidespread portability, using the Cetus framework, which has been extended to

3.6 Summary 93

support Fortran 77 in addition to C codes. The compiler performs all necessaryanalyses and transformations to the code, including static data ow, communica-tion analyses and the insertion of checkpoints at key places in the code. All theseanalyses have been thoroughly described, as well as their implementation detailsfocusing on the design decisions taken to enable the reusability of the platform. Byusing a semantic catalog, the compiler is able to analyze a variety of APIs for com-munications, le I/O, etc., and thus it is not tied to any specic interface such asMPI. Concluding the chapter, the related work has been covered, giving an overviewof other existing techniques for checkpoint insertion and communication analysis.

Chapter 4

Experimental Results

4.1. Introduction

In order to conduct the experimental evaluation of the CPPC checkpointingframework, two dierent sets of experiments were designed and executed to sepa-rately assess the performance of both the CPPC compiler and the CPPC library.Twelve applications were selected for testing, separated into three dierent types.First, the eight applications in the NPB-MPI v3.1 benchmarks [3]. These haveshort runtimes, which makes them ill-suited choices for checkpoint insertion. How-ever, they are well-known and widespread, which makes them a good choice forcomparison purposes. The next two programs are scientic applications in usein the Galician Supercomputing Center (CESGA), called CalcuNetw and Fekete.While these applications do not have particularly long runs, they are good choicesfor evaluating the performance of the framework in applications currently in usein supercomputing centers. Finally, two large-scale applications called DBEM andSTEM were also added to test the tool in long running programs. Table 4.1 givesmore details about the test applications used.

This chapter is divided into two sections. Section 4.2 covers experimental resultsfor the CPPC compiler. Section 4.3 details execution tests for the CPPC library.

95

96 Chapter 4. Experimental Results

Table4.1:

Summaryof

testapplications

Source

Application

Description

NASNPB-M

PIv3.1

BT

Asimulated

CDFapplicationthat

uses

animplicitalgorithm

tosolve3-

dimensional

compressibleNavier-Stokes

equations.

The

nitedierences

solution

totheproblem

isbasedon

anAlternate

Direction

Implicit(A

DI)

approximatefactorizationthat

decouplesthe

x,yand

zdimensions.The

resultingsystem

sareBlock-T

ridiagonal

of5x

5blocks

andaresolved

se-

quentially

alongeach

dimension.

CG

UsesaConjugateGradientmethodto

compute

anapproximationto

the

smallesteigenvalue

ofalarge,sparse,unstructured

matrix.

EP

EmbarrassinglyParallelbenchmark.

Itgeneratespairsof

Gaussianran-

dom

deviates

accordingto

aspecicscheme.

FT

Containsthecomputational

kernel

ofafastFourierTransform

(FFT)-

basedspectral

method.

Itperform

sthree1-dimensional

FFT's,onefor

each

dimension.

ISIntegerSort:Itworks

withalistofsm

allintegervalues,notreallysorting

them

butassigningeverylistmem

ber

anumber

indicating

thepositionin

thesorted

list.

LU

Asimulated

CFD

applicationthat

uses

asymmetricsuccessive

over-

relaxation

(SSO

R)methodto

solveaseven-block-diagonalsystem

result-

ingfrom

nite-dierence

discretization

ofthe3-dimensionalNavier-Stokes

equationsby

splitting

itinto

blockLow

erandUpp

ertriangular

system

s.

MG

UsesaV-cycleMultiGridto

compute

thesolution

ofthe3-dimensional

scalar

Poisson

equation.The

algorithm

works

continuously

onasetof

gridsthat

aremadebetweencoarse

andne.

4.1 Introduction 97Table4.1:

Summaryof

testapplications

(continued)

Source

Application

Description

NASNPB-M

PIv3.1

SP

Asimulated

CFD

applicationthat

hasasimilarstructureto

BT.The

nitedierencessolution

totheproblem

isbasedon

aBeam-W

arming

approximatefactorizationthat

decouplesthe

x,yand

zdimensions.The

resultingsystem

hasScalarPentadiagonalbandsof

linearequationsthat

aresolved

sequentially

alongeach

dimension.

CESG

ACalcuNetw[43]

Calculatessomecharacterization

measurements

inagivennetwork,

con-

sistingof

asetof

nodesor

vertices

joined

together

inpairsby

links

oredges,andcomparesitwithanumber

ofrandom

networks

specied

bytheuser.The

program

calculates

thesubgraph

centrality(SC),SC

odd,

SCeven,bipartivity,networkcommunicability,andnetworkcommunica-

bilityforconnectednodes.

Fekete

[17]

Determines

thepositionof

acertainnumber

ofpointson

a2-dimensional

sphere

such

that

thepotentialenergy

produced

bytheinteractionofthese

pointsisminimum

.Thisisthe7thof

theSm

ale'sproblems[60].

DBEM

[30]

Crack

grow

thanalysisusingtheDualBoundaryElementMethod.

The

analysis

leadsto

alargenumber

ofdiscretizedequationsthat

grow

ateverystep

whenthecrackgrow

this

evaluated.

Itsolves

theresulting

denselinearsystem

usingtheGMRESiterativemethod,

regarded

asthe

mostrobustof

theKrylovsubspace

iterativemethods.

STEM-II[42]

Usedto

know

inadvancehowthemeteorologicalconditions,obtained

from

ameteorogicalprediction

model,would

aecttheem

issionsof

pollutants

bythepow

erplantof

AsPontes(A

Coruña,Spain)

inorderto

fulllEU

regulations.

The

underlying

equation

isatime-dependent,3-dimensional

partialdierentialatmosferic-diusionequation.

98 Chapter 4. Experimental Results

4.2. Compiler

This section focuses on two main issues when using the CPPC compiler to instru-ment parallel applications: correction of the resulting code, covered in Section 4.2.1;and the performance obtained when executing the necessary analyses and transfor-mations, detailed in Section 4.2.2.

4.2.1. Analysis of the instrumented codes

In this section, the results of the checkpoint insertion analyses are shown. Thefollowing subsections detail the results of the rst and second thresholding steps ofthe checkpoint insertion algorithm for all the test applications, and compare themto the optimal manual checkpoint placement performed by the authors during theinitial tests of the applications. While some extra checkpoints are inserted in appli-cations with similar heuristic values for both time-consuming loops and negligibleones, the compiler never leaves a time-consuming loop without checkpointing. Thismeans that the analysis behaves in a conservative way. Thus, while the resultingcode may not be completely optimal, application progress is guaranteed.

Regarding the communication matching, the compiler correctly detects safe pointsfor all test applications.

• NAS BT

Figure 4.1(a) shows the heuristic h(l) as calculated for the NAS BT application.The gure is the same as the one shown for exemplifying the checkpoint insertionalgorithm described in Section 3.2.7. The rst step of the thresholding processselects a subset of four loop nests as candidates for checkpoint insertion. The valuesof h(l) for these loops are detailed in Table 4.2. Loop #1 corresponds to the maincomputational loop in BT, and is the place where a manual checkpoint was insertedby the authors in the preliminary application tests. The second thresholding step,shown in Figure 4.1(b), determines that the loops in C0 are responsible for the56.94% of the total variation of h(l) in the H subset, and therefore selects onlyloop #1 for checkpoint insertion, which matches the optimal checkpoint insertion inthe manual analysis.

4.2 Compiler 99

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.1: Checkpoint insertion for NAS BT

100 Chapter 4. Experimental Results

Table 4.2: Detail of loops selected by the shape-based threshold for NAS BTLoop # File Line Statements Accesses h(l)

1 bt.f 179 2075 5547 0.77552 exact_rhs.f 24 107 490 3.11493 initialize.f 44 30 132 4.23684 error.f 25 19 47 4.8837

Total program statements 5084Total program accesses 13502

• NAS CG

Figure 4.2(a) shows the heuristic h(l) calculated for the NAS CG application.The rst step of the thresholding process selects a subset of ve loop nests as can-didates for checkpoint insertion. The values of h(l) for these loops are detailed inTable 4.3. Loop #1 corresponds to the main computational loop in CG, and is theplace where a manual checkpoint was inserted by the authors in the preliminaryapplication tests. The second thresholding step, shown in Figure 4.2(b), createstwo dierent clusters, and selects C0 for checkpointing, responsible for 59.16% ofthe total variation of h(l) in the H subset. Two loops are included in C0, loop #2being a exact copy of loop #1. The only dierence is that it performs a singleuntimed iteration which initializes all code and data page tables, and that it doesnot perform the conditional statement found in line 500 in le cg.f. This accountsfor the dierence in h(l) for both loops, which is negligible as shown in the tables.After executing this single-iteration loop, the application reinits and starts timing.The compiler was not programmed to detect single-iteration loops as not real loops,and therefore this loop is selected as well for checkpointing, which may create anadditional redundant checkpoint depending on the runtime conguration of the li-brary, particularly on whether or not the user decides to activate checkpointing eachtime a new checkpoint is reached, regardless of the number of calls performed to thecheckpoint function.

• NAS EP

Figure 4.3(a) shows the heuristic h(l) as calculated for the NAS EP application.The rst step of the thresholding process selects a subset of two loop nests as can-didates for checkpoint insertion. The values of h(l) for these loops are detailed in

4.2 Compiler 101

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.2: Checkpoint insertion for NAS CG

102 Chapter 4. Experimental Results

Table 4.3: Detail of loops selected by the shape-based threshold for NAS CGLoop # File Line Statements Accesses h(l)

1 cg.f 441 97 220 1.02202 cg.f 344 96 217 1.03293 cg.f 1453 28 20 2.61604 cg.f 1606 14 10 3.17235 cg.f 910 4 10 3.7163

Total program statements 354Total program accesses 647

Table 4.4: Detail of loops selected by the shape-based threshold for NAS EPLoop # File Line Statements Accesses h(l)

1 ep.f 189 16 23 1.32852 ep.f 165 2 2 3.2923

Total program statements 70Total program accesses 114

Table 4.4. Loop #1 corresponds to the main computational loop in EP, and is theplace where a manual checkpoint was inserted by the authors in the preliminaryapplication tests. The second thresholding step, shown in Figure 4.3(b), assignseach point in H to a dierent cluster, and selects loop #1 for checkpoint insertion,responsible for 100% of the total variation of h(l) in the H subset. Note that anysituation in which only two points reach the cluster-based thresholding step willresult in the algorithm selecting only the one with the bigger h(l) value. This makesperfect sense, since the shape-based threshold would not select only two loop nestsif they had similar heuristic values.

• NAS FT

Figure 4.4(a) shows the heuristic h(l) as calculated for the NAS FT application,which is larger and structurally more complex than the two previous ones. The rststep of the thresholding process selects a subset of six loop nests as candidates forcheckpoint insertion. The values of h(l) for these loops are detailed in Table 4.5.Loop #1 corresponds to the main computational loop in FT, and is the place wherea manual checkpoint was inserted by the authors in the preliminary applicationtests. The second thresholding step, shown in Figure 4.4(b), creates three dierent

4.2 Compiler 103

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.3: Checkpoint insertion for NAS EP

104 Chapter 4. Experimental Results

Table 4.5: Detail of loops selected by the shape-based threshold for NAS FTLoop # File Line Statements Accesses h(l)

1 ft.f 159 189 337 1.20052 ft.f 676 9 24 3.53983 ft.f 689 9 24 3.53984 ft.f 702 9 24 3.53985 ft.f 1016 7 9 4.07506 ft.f 271 3 7 4.5521

Total program statements 743Total program accesses 1261

clusters, and selects only loop #1 for checkpoint insertion as responsible for 69.80%

of the total variation of h(l) in the H subset, which matches the optimal checkpointinsertion in the manual analysis.

• NAS IS

Figure 4.5(a) shows the heuristic h(l) as calculated for the NAS IS application,the only of the NPB ones written in C. As can be seen, the shape of h(l) is quiteatypical, due to the fact that this application is very small and all loops have similarsizes. In fact, this is the application with the smallest total variation in h(l), andthe one in which h(l) most closely resembles a straight line. This causes the rststep of the thresholding process to select a high number of loop nests as candidatesfor checkpoint insertion, ve out of the total six in IS. The values of h(l) for theseloops are detailed in Table 4.6. Loop #2 corresponds to the main computationalloop, and the only one manually selected in the preliminary application tests. Thesecond thresholding step, shown in Figure 4.5(b), creates three dierent clusters. Itdoes not select C0 only, since it is only responsible for 41.41% of the total variationof h(l) in the H subset. Instead, it selects both C0 and C1, which are togetherresponsible for 61.94% of such variation.

Although the checkpoint insertion algorithm does indeed select the main compu-tational loop for checkpoint insertion, thus guaranteeing execution progress in thepresence of failures, two other loops are conservatively selected in addition to it,making it a non-optimal choice.

4.2 Compiler 105

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.4: Checkpoint insertion for NAS FT

106 Chapter 4. Experimental Results

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.5: Checkpoint insertion for NAS IS

4.2 Compiler 107

Table 4.6: Detail of loops selected by the shape-based threshold for NAS ISLoop # File Line Statements Accesses h(l)

1 is.c 425 90 154 0.65702 is.c 976 56 39 1.45953 is.c 396 36 59 1.47164 is.c 387 23 38 1.85735 is.c 882 16 10 2.5947

Total program statements 242Total program accesses 260

Table 4.7: Detail of loops selected by the shape-based threshold for NAS LULoop # File Line Statements Accesses h(l)

1 ssor.f 78 452 2251 1.04662 erhs.f 383 43 151 3.48283 erhs.f 118 33 116 3.7122

Total program statements 1961Total program accesses 7415

• NAS LU

Figure 4.6(a) shows the heuristic h(l) as calculated for the NAS LU application.The rst step of the thresholding process selects a subset of three loop nests ascandidates for checkpoint insertion. The values of h(l) for these loops are detailedin Table 4.7. Loop #1 corresponds to the main computational loop in LU, and isthe place where a manual checkpoint was inserted by the authors in the preliminaryapplication tests. The second thresholding step, shown in Figure 4.6(b), creates twoclusters, and selects only loop #1 for checkpoint insertion as responsible for 91.39%

of the total variation of h(l) in the H subset, which matches the optimal checkpointinsertion in the manual analysis.

• NAS MG

Figure 4.7(a) shows the heuristic h(l) as calculated for the NAS MG application.The rst step of the thresholding process selects a subset of two loop nests as can-didates for checkpoint insertion. The values of h(l) for these loops are detailed inTable 4.8. Loop #1 corresponds to the main computational loop in MG, and is the

108 Chapter 4. Experimental Results

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.6: Checkpoint insertion for NAS LU

4.2 Compiler 109

Table 4.8: Detail of loops selected by the shape-based threshold for NAS MGLoop # File Line Statements Accesses h(l)

1 mg.f 245 503 1354 0.97072 mg.f 2100 27 46 3.5997

Total program statements 1597Total program accesses 3862

Table 4.9: Detail of loops selected by the shape-based threshold for NAS SPLoop # File Line Statements Accesses h(l)

1 sp.f 150 927 1783 1.00772 exact_rhs.f 23 107 490 2.50443 initialize.f 45 30 132 3.62624 error.f 26 19 47 4.27315 initialize.f 103 12 40 4.5427

Total program statements 2801Total program accesses 6009

place where a manual checkpoint was inserted by the authors in the preliminary ap-plication tests. As in the NAS EP application, the second thresholding step, shownin Figure 4.7(b) creates two clusters, and selects only one loop for checkpointingwhich is responsible for 100% of the variation of h(l) in the H subset. This behaviormatches the optimal checkpoint insertion in the manual analysis.

• NAS SP

The structure of the NAS SP application is very similar to that of the NAS BT,thus the outcome of the checkpoint insertion analysis is very similar to the outcomefor the former. The rst and second thresholding steps are shown in Figures 4.8(a)and 4.8(b), respectively. Values for h(l) in the H subset, which contains ve loops,are detailed in Table 4.9. The cluster-based thresholding step creates three clusters.C0 is only responsible for a 42.34% of the total variation of h(l) in the H subset, soboth C0 and C1 get checkpointed, accounting for a 92.37% of the variation of h(l).Manual checkpointing insertion is limited to the loop nest in C0, which means thatall three loops in C1 are conservatively checkpointed.

110 Chapter 4. Experimental Results

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.7: Checkpoint insertion for NAS MG

4.2 Compiler 111

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.8: Checkpoint insertion for NAS SP

112 Chapter 4. Experimental Results

Table 4.10: Detail of loops selected by the shape-based threshold for CESGA Cal-cuNetw

Loop # Statements Accesses h(l)

1 162 180 0.75482 28 14 2.62633 14 11 3.03214 9 3 3.7883

Total program statements 392Total program accesses 423

• CESGA CalcuNetw

The code of the following applications is not in the public domain, and thereforeit makes no sense to give details about code locations of the selected loops. Heuristicvalues for loop nests in the H subset, however, are included in Table 4.10 for com-pletitude. It is interesting to detail the shape of h(l) and how the insertion algorithmbehaves, since these are naturally better candidates for checkpoint insertion thanthe NAS Parallel Benchmarks. Figure 4.9(a) shows the rst step of the thresholdingalgorithm, while Figure 4.9(b) depicts the second one. As can be seen, the shape ofthe application is similar to the larger ones in the NPB set. The shape-based thresh-old selects a subset of four loop nests as candidates for checkpoint insertion, whichare categorized in three dierent clusters by the second thresholding step. Loop #1is then selected for checkpoint insertion as it contains 61.70% of the variation of h(l)

in the H subset. This loop is the main computational loop in the application, andthe decision matches the manual one performed during the initial assessment of theapplication.

• CESGA Fekete

Again, the automatic analysis for this application matches the outcome of themanual decision process. The rst and second thresholding steps are shown inFigures 4.10(a) and 4.10(b), respectively. Values for h(l) in the H subset are detailedin Table 4.11. Two clusters are created in the second thresholding step. Loop #1in C0 is the only nest to be checkpointed, being responsible for 68.84% of the totalvariation of h(l) in the H subset.

4.2 Compiler 113

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.9: Checkpoint insertion for CESGA CalcuNetw

114 Chapter 4. Experimental Results

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.10: Checkpoint insertion for CESGA Fekete

4.2 Compiler 115

Table 4.11: Detail of loops selected by the shape-based threshold for CESGA FeketeLoop # Statements Accesses h(l)

1 77 181 0.30772 9 9 2.52883 2 4 3.5342

Total program statements 119Total program accesses 238

• DBEM

This is the biggest application in the experimental set, with a total of 92 loopnests being considered by the compiler. Out of them, 10 are selected by the shape-based threshold, as shown in Figure 4.11(a), and categorized into ve dierent clus-ters by the second thresholding method detailed in Figure 4.11(b). Values for h(l)

in the H subset are detailed in Table 4.12. Loop #1 is responsible for the creationof the equation systems that are resolved later in loop #2, which is the main com-putational loop of the application. The manual analysis performed in the initialassessment of the application determined that both loops should be checkpointed.The automatic analysis decides to checkpoint both, and also two more loops that areconservatively selected. Although loop #2 is considerably bigger than both of them,the huge size of loop #1 makes them comparable when in the context of the entireapplication. Together, C0 and C1 are responsible for 65.69% of the total variationof h(l) in H.

• STEM-II

This application, the second biggest in the experimental set, has most of its codeinside a single loop, which is the one manually selected for checkpointing during theinitial tests and also the only one checkpointed by the automatic checkpoint insertionalgorithm. Figures 4.12(a) and 4.12(b) show the rst and second thresholding steps,respectively. As can be seen in Table 4.13, loop #1 is so big that it completelyneglects the rest of the nests in the application, and is the only selected from thetwo belonging to the H subset.

116 Chapter 4. Experimental Results

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.11: Checkpoint insertion for DBEM

4.2 Compiler 117

(a) First step: shape-based threshold

(b) Second step: cluster-based threshold

Figure 4.12: Checkpoint insertion for STEM-II

118 Chapter 4. Experimental Results

Table 4.12: Detail of loops selected by the shape-based threshold for DBEMLoop # Statements Accesses h(l)

1 1924 2869 0.50172 333 438 2.07973 214 426 2.28384 182 457 2.32365 135 237 2.73856 84 180 3.06407 89 66 3.47478 52 95 3.54999 60 69 3.626610 53 41 3.9065

Total program statements 3330Total program accesses 5262

Table 4.13: Detail of loops selected by the shape-based threshold for STEM-IILoop # Statements Accesses h(l)

1 1673 4305 0.32972 9 12 5.1538

Total program statements 2731Total program accesses 5635

4.2.2. Compilation times

A summary of the characteristics of the test applications, including number ofles, lines of code (LOCs), time it takes the CPPC compiler to instrument them,number of loop nests in the code, and number of checkpoints and variable registra-tions inserted by the CPPC compiler is shown in Table 4.14. Compilation timeswere measured in a desktop machine, with an Intel Core2 Duo CPU at 3.00 Ghz.and 1 GB of RAM. Although the CPPC compiler is a experimental tool and is notoptimized for production use, it can be seen that compile times are acceptable forall test applications, and mostly dependent on the number of lines of code of thesource. The higher compile time is obtained for the DBEM application, with 12533LOCs, and is 77.14 seconds.

A more in-depth analysis of compilation times is given in Table 4.15, where theyhave been broken down into the times for each of the analyses described in Chap-ter 3. Times for code parsing, insertion of initialization and shutdown CPPC calls,

4.3 Library 119

instrumentation of non-portable calls and open les, conversion to CPPC state-ments, data ow analysis, communication analysis, checkpoint insertion, language-dependent transforms and code generation are detailed. Although the number ofapplications is not nearly enough as to develop a complete mathematical model ofexecution times, some tendencies can be inferred from these times. Most of the anal-yses run in linear time with respect to the lines of code of the compiled application.Particularly, code parsing, the insertion of the CPPC initialization and shutdowncalls, the analysis of non-portable calls, instrumentation of open les, transforma-tion to CPPC statements, and language-dependent transforms appear to be O(n),being n the number of LOCs in the code. The data ow analyses are O(n) as well,but very dependent on the programming language the application is coded in. Ifa linear regression model were tted to the times for the data ow analysis, twodierent models would have to be created for the analysis of C and Fortran 77 ap-plications. The communications analysis tends to O(n2). The checkpoint insertionanalysis does not depend on the LOCs of the application, but rather on the numberof loop nests (#L), being O(#L2). As for the times described as code generation,they include many compilation passes, such as insertion of the actual registrationcalls for restart-relevant variables, control ow code, etc. Thus, it cannot be an-alyzed precisely as a whole, but should be broken down in each of the individualanalysis that are performed. However, as a general pattern, it may be characterizedas O(n2).

4.3. Library

Runtime tests of the CPPC library were performed on the Finis Terrae super-computer hosted by CESGA (Galician Supercomputing Center). It consists of 144computation nodes:

142 HP Integrity rx7640 nodes with 16 Itanium Montvale cores and 128 GBof RAM per node.

1 HP Integrity Superdome node, with 128 Itanium Montvale cores and 1024GB of RAM.

1 HP Integrity Superdome node, with 128 Itanium 2 cores and 384 GB ofRAM.

120 Chapter 4. Experimental Results

Table

4.14:Statistics

andcom

pilationtim

esfor

testapplications

Applica

tion

Files

LOCs

Compile

time

Loopnests

#Chkpts.

#Registe

rs

NAS NPB-MPI v3.1

BT

183650

14.19s.

251

121CG

11044

2.61s.

132

38EP

1180

0.90s.

41

14FT

11269

4.82s.

201

31IS

1672

3.88s.

63

34LU

253086

7.78s.

351

74MG

11618

13.08s.

121

34SP

243148

13.69s.

254

143

CESG

ACalcuN

etw5

81057.34

s.14

128

Fekete1

1820.86

s.6

117

DBEM

4212533

77.14s.

924

279ST

EM

1106506

15.14s.

241

94

4.3 Library 121Table4.15:Breakdownof

compilation

times

fortestapplications

Application

Totaltime

Parsing

Init+Shutdown

Non-portable

Openles

CPPCstmts.

NASNPB-MPIv3.1

BT

14.19s.

2974

ms.

346ms.

106ms.

110ms.

445ms.

CG

2.61

s.415ms.

80ms.

15ms.

17ms.

84ms.

EP

0,9s.

185ms.

55ms.

6ms.

6ms.

14ms.

FT

4.82

s.992ms.

159ms.

43ms.

48ms.

264ms.

IS3.88

s.395ms.

77ms.

22ms.

23ms.

22ms.

LU

7.78

s.1564

ms.

203ms.

57ms.

61ms.

282ms.

MG

13.08s.

1278

ms.

217ms.

60ms.

64ms.

332ms.

SP13.69s.

1815

ms.

199ms.

57ms.

60ms.

230ms.

CESG

ACalcunetw

24.55s.

906ms.

110ms.

96ms.

104ms.

130ms.

Fekete

0.86

s.189ms.

54ms.

4ms.

6ms.

12ms.

DBEM

77.14s.

4975

ms.

344ms.

117ms.

144ms.

443ms.

STEM-II

15.14s.

1786

ms.

220ms.

65ms.

82ms.

195ms.

Application

Data

ow

Comms.

Chkpt.insertion

Language

Codegeneration

NASNPB-MPIv3.1

BT

3170

ms.

2449

ms.

851ms.

100ms.

2851

ms.

CG

519ms.

636ms.

80ms.

18ms.

515ms.

EP

123ms.

78ms.

11ms.

4ms.

229ms.

FT

1180

ms.

837ms.

273ms.

46ms.

952ms.

IS1043

ms.

1679

ms.

48ms.

19ms.

337ms.

LU

1920

ms.

1304

ms.

388ms.

55ms.

1565

ms.

MG

1561

ms.

7707

ms.

317ms.

60ms.

1349

ms.

SP2211

ms.

1756

ms.

783ms.

52ms.

6239

ms.

CESG

ACalcunetw

21955ms.

0ms.

198ms.

84ms.

999ms.

Fekete

146ms.

70ms.

15ms.

4ms.

176ms.

DBEM

17198ms.

15233ms.

20334ms.

101ms.

18267ms.

STEM-II

5946

ms.

3553

ms.

1448

ms.

48ms.

1387

ms.

122 Chapter 4. Experimental Results

Table 4.16: Number of nodes and cores used for runtime tests for each applicationApplication Nodes Cores per node Total cores

NASNPB-M

PIv3.1 BT 3 12 36

CG

2 16 32

EPFTISLUMGSP 3 12 36

CESGACalcuNetw 1 1 1Fekete

2 16 32DBEMSTEM-II

The nodes are conected through an Inniband 4x DDR network with a band-with of 20 Gbps. The applications were executed on the HP Integrity rx7640 nodes.Table 4.16 details the number of nodes used for each application in runtime tests.Whenever possible, two nodes with 16 cores each were used for the execution. How-ever, the NAS BT and SP applications require the number of parallel processes tobe a square, which caused the number of processes to be raised to 36, allocating 3nodes and using 12 cores of each of them. The remaining 4 cores in each node wereallocated to prevent applications belonging to other users from being executed onthem, thus disrupting the results. CESGA CalcuNetw is a sequential application.A whole node was allocated for its execution, while only one of its cores was usedfor the tests.

Executed measurements include: generated state le sizes, time for state le gen-eration, checkpointing overhead and restart times. For proving portability, cross-restarts were executed on the Finis Terrae using state les generated by the Muxiacluster, hosted by the Computer Architecture Group of the University of A Coruña,consisting of Intel Xeon 1.8 Ghz nodes, with dierent compilers and MPI implemen-tations than those found in the Finis Terrae.

The NAS CG, IS and SP, and DBEM applications create more than one check-point le during their execution. In order to provide normalized results for thestate le sizes, state le creation time and restart time tests, these parameters weremeasured for the biggest checkpoints created in each case. This makes it easier toprovide graphical comparisons of the results for dierent applications.

4.3 Library 123

4.3.1. State le sizes

When using CPPC's spatially coordinated technique, the incurred overhead willonly depend on the overhead introduced by the checkpoint le dumping. Thisoverhead heavily depends on the size of the data to be dumped. Thus, the rstparameter to be measured is how the variable level approach aects checkpointle sizes. In order to analyze this eect, state le sizes have been measured fordierent checkpointing congurations, and compared to sizes obtained using a fullcheckpointer. For most applications, le sizes also vary depending on the number ofprocesses involved in the execution, because the sizes of array data are modied tomatch the problem size. The exceptions are NAS EP, CESGA Fekete, DBEM andSTEM-II, which statically allocate a xed array size that determines the maximumallowed problem size. The results of this test are shown in Figures 4.13 to 4.23.The values tagged as Automatic are le sizes obtained by the automatic variableregistration included in the compiler. Compressed shows le sizes using the HDF-5writer with compression enabled. Compression is only applied to vector variables(pointers and arrays) larger than a certain user-specied limit, which in this casewas set to 2000 elements. For comparison purposes, Optimal shows the optimal lesizes obtained by a manual analysis, and Full data presents the sizes obtained fora checkpointer that stores all the application data. Figure 4.24 shows a summary ofthe sizes obtained for all applications using the number of processes used for runtimetests shown in Table 4.16, and includes the CESGA CalcuNetw application, whichdoes not have a gure of its own because it is a sequential application.

Sizes obtained using automatic analyses are close to the optimal ones, for mostapplications. The dierence between both is due to the registration of unnecessaryarray sections because of the conservative approach of the compiler, as explainedin Section 3.2.5. Variable level checkpointing usually achieves very important sizereductions when compared to full data sizes. Table 4.17 gives the exact details onreduction percentages for the number of processes used in runtime tests for eachapplication. It includes the size of the les created by the automatic live variableanalysis, the reduction obtained with respect to the size of the les created by a fulldata checkpointer, and the optimal reduction (bracketed).

The high compression rates obtained for DBEM and STEM-II (97.86% and96.10%, respectively) are due to the fact that these applications statically allocatearrays which are oversized to t a maximum problem size. This is directly relatedwith the fact that checkpoint sizes for these applications do not vary with the num-

124 Chapter 4. Experimental Results

Table 4.17: Performance of the automatic variable registration algorithmApplication Automatic size (MB) Reduction (Optimal)

NASNPB-M

PIv3.1 BT 17.30 93.77% (98.55%)

CG 19.77 93.55% (93.64%)EP 1.04 88.87% (99.71%)FT 40.10 87.32% (92.38%)IS 72.1 78.28% (92.76%)LU 8.26 51.81% (64.26%)MG 15.64 60.49% (63.15%)SP 19.77 86.05% (91.99%)

CESGAFekete 1.60 82.59% (92.81%)

CalcuNetw 3.12 73.78% (73.80%)DBEM 275.67 5.94% (10.70%)STEM-II 114.13 39.72% (83.42%)

ber of processes executing them. Rather, the more processes, the bigger the problemthat can be solved. As a result, an important amount of empty memory is allocatedin our tests, resulting in high compression rates.

4.3.2. State le creation time

The Finis Terrae machine exhibits a high variability between times obtainedby dierent executions of the same experiment. Allocating entire nodes reducesthe variability, but does not entirely solve the problem. Thus, performing a largenumber of experiments and statistically analyzing the results is a better approachthan just giving the minimum, maximum, or mean times for each experiment. Theexperimental approach consisted on measuring the creation time for the bigger statele for each application a minimum of 500 times, followed by the elimination ofoutliers and the calculation of the 99% condence interval for the expected creationtimes. The approach used for outlier identication was to discard observationsbigger than a certain threshold. This threshold is dened for each application asthe third quartile of the data series plus 1.5 times the interquartile range (IQR).This is a classical approach to outlier identication [68]. Table 4.18 shows the 99%condence interval for the mean creation times of a standard HDF-5 state le, thesame HDF-5 le including a CRC-32 error detection scheme, and for compressedHDF-5 les for all applications. Also, the minimum obtained times are shown forcomparison purposes. Note that these times correspond to the raw dumping time

4.3 Library 125

1 4 9 16 25 36 49 640

100

200

300

400

500

600

700

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.13: File sizes for NAS BT

1 2 4 8 16 32 640

200

400

600

800

1000

1200

1400

1600

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.14: File sizes for NAS CG

126 Chapter 4. Experimental Results

1 2 4 8 16 32 640

2

4

6

8

10

12

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.15: File sizes for NAS EP

1 2 4 8 16 32 640

500

1000

1500

2000

2500

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.16: File sizes for NAS FT

4.3 Library 127

1 2 4 8 16 32 640

500

1000

1500

2000

2500

3000

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.17: File sizes for NAS IS

1 2 4 8 16 32 640

50

100

150

200

250

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.18: File sizes for NAS LU

128 Chapter 4. Experimental Results

1 2 4 8 16 32 640

50

100

150

200

250

300

350

400

450

500

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.19: File sizes for NAS MG

1 4 9 16 25 36 49 640

50

100

150

200

250

300

350

400

450

500

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.20: File sizes for NAS SP

4.3 Library 129

1 2 4 8 16 32 640

2

4

6

8

10

12

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.21: File sizes for CESGA Fekete

1 2 4 8 16 32 640

50

100

150

200

250

300

350

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.22: File sizes for DBEM

130 Chapter 4. Experimental Results

1 2 4 8 16 32 640

20

40

60

80

100

120

140

160

180

200

Optimal Automatic Compressed Full data

# processes

sta

te fi

le s

ize

(M

B)

Figure 4.23: File sizes for STEM-II

BT CG EP FT IS LU MG SP Fekete CalcuNetw DBEM STEM0

50

100

150

200

250

300

350

Optimal Automatic Compressed Full data

sta

te fi

le s

ize

(M

B)

Figure 4.24: Summary of le sizes

4.3 Library 131

of a single checkpoint, not the real contribution of dumping times to the checkpointoverhead, which is reduced by using multithreaded dumping.

Written data are tagged by the HDF-5 library to allow for conversions, if needed,when restarting the application. This improves checkpoint mode performance, mov-ing the conversion overhead to the restart mode, which should be a much less fre-quent operation. A regression analysis on the dumping times with respect to thecheckpoint le sizes yields a linear relationship between both factors in both thestandard and the CRC-32 le creations. Using compression, however, the relation-ship is not linear, since the dumping time depends on the number of variables tobe compressed and the entropy of the contained data. As can be seen, compressingdata heavily increases overall dumping times. Therefore, it should be enabled onlywhen the physical size of state les is critical: for instance, if there are problemswith disk quotas or when the les are going to be transferred using a slow net-work. Also, it should be noted that the interval obtained for the mean creationtimes of compressed les for the NAS MG application is particularly large (about0.25 seconds). This is due to the fact that the entropy of the stored data does notremain constant through the execution, causing compression times to increase asthe execution progresses, and a high variance of the obtained times.

Figure 4.25 graphically compares the results shown in the table above, using themaximum estimated value (i.e. the upper end of the condence intervals) for eachapplication.

4.3.3. Checkpoint overhead

To reduce the overhead introduced by le generation, multihreaded state dump-ing has been implemented. When performing the overhead tests, the problem ofthe experimental variability in Finis Terrae has to be carefully dealt with. The loadof the allocated nodes dramatically impacts execution times. To handle this situa-tion, full nodes were reserved for the execution of the experimental applications asexplained before. Moreover, it is important that the applications are executed onthe same nodes and in the same conditions. Thus, once the nodes were allocated, asequence of N experiments was run on them randomly selecting the type of each ofthem before being executed. Thus, for instance, if 500 experiments were scheduled,before starting each of them a random number was generated, and depending on itbeing odd or even the original version of the application or the CPPC-instrumentedone was run. For the shorter applications, the number N of experiments was limited

132 Chapter 4. Experimental Results

Table

4.18:Checkp

ointle

creationtim

es(seconds)

STANDARD

CRC-32

COMPRESSED

Applica

tion

xm

inx

max

min

xm

inx

max

min

xm

inx

max

min

NAS NPB-MPI v3.1

BT

0.13740.1436

0.11620.2221

0.22860.2016

1.32481.3299

1.3095CG

0.09040.0991

0.08960.1914

0.19880.1728

1.72161.7256

1.7093EP

0.00650.0066

0.00610.0112

0.01130.0108

0.11950.1198

0.01187FT

0.16120.1717

0.12590.3601

0.36850.3217

3.42363.4339

3.3957IS

0.28750.2996

0.24570.6500

0.65990.6008

6.27446.3449

6.1869LU

0.07760.0810

0.06610.1145

0.11720.1040

0.61910.6279

0.6006MG

0.10350.1077

0.08900.1796

0.18400.1666

0.67950.9111

0.5113SP

0.16330.1692

0.14710.2657

0.27060.2509

1.26291.2699

1.2387

CESG

AFekete

0.02510.0275

0.01870.0309

0.03220.0267

0.08180.0824

0.0802CalcuN

etw0.0207

0.02140.0189

0.03530.0363

0.03130.0964

0.09680.0951

DBEM

2.23842.2725

2.13393.6321

3.67693.4345

7.02847.0646

6.8588ST

EM-II

0.98981.0147

0.87241.5406

1.56171.4581

4.30674.3527

4.0857

4.3 Library 133

BT CG EP FT IS LU MG SP Fekete CalcuNetw DBEM STEM-II0

1

2

3

4

5

6

7

8

Standard CRC-32 Compressed

time

(s)

Figure 4.25: Maximum mean dumping times for test applications

to 500, a number which should provide statistically representative results withoutthe need for more repetitions. For the larger ones this number was determined bythe maximum allocation time of the nodes, which was 10 hours for the number ofallocated cores. Performing more than one allocation is not a valid approach, sinceusually the obtained time series may not be assumed to come from the same statis-tical distribution. This experimental setup is designed to ensure that the variabilityexhibited by the machine aects all types of experiments to the same degree. Out-liers are identied, as in the previous experiment, by removing observations biggerthan the third quartile plus 1.5 times IQR in each of the series. Table 4.19 showsthe number of regular and CPPC runs performed for each application, a 99% con-dence interval for the checkpointing overhead, the minimum execution time forboth the regular and CPPC versions of the code, and the maximum overhead per-centage calculated as the upper limit of the overhead condence interval divided bythe minimum original runtime. As can be seen, sometimes the minimum execution

134 Chapter 4. Experimental Results

time for the CPPC version is below the minimum time for the original version ofthe application. This evidences the need for statistical analyses of execution times.Note that, even when considering the maximum possible overhead at 99% condenceand comparing it to the minimum time obtained for the execution of the originalcode, the overhead percentages remain low, usually in the 1% range, except for theIS application, which runs for about 25 seconds and therefore the 9.37% overheadobtained does not account for more than 2.5 seconds. These overheads include,besides le generation, all CPPC instrumentation (e.g. variable and parameter reg-istration). Files were generated using the HDF-5 writer, with CRC-32 and withoutcompression. One state le per checkpoint was generated for all the applications inthe table.

The previously detailed experimental setup is not viable for the long-runningapplications, DBEM and STEM-II, since they run for more than 6 and 23 hoursrespectively. In order to obtain statistically signicant measurements for these ap-plications in the same way as for the shorter ones would require years of computationtime. Instead, these applications were run in the Muxia cluster, owned by the Com-puter Architecture Group of the University of A Coruña, formed by Intel Xeon 1.8Ghz nodes with 1 GB of RAM, connected through an SCI network. This clusterdoes not exhibit the high loads nor the high variability experienced in the FinisTerrae, and therefore makes it possible to reach plausible conclusions without theneed for a large number of experiments. Thus, 10 executions of each version of thecodes were run. Table 4.20 details the minimum execution times for the originalversions of both DBEM and STEM-II executed on 4 cores, the checkpoint overheadincurred in seconds, and the overhead percentage. The checkpointing frequency forthese applications was adjusted to approximately one checkpoint per hour, meaningthat 23 checkpoints were created during the DBEM executions, and 7 checkpointsduring the STEM-II ones.

4.3.4. Restart times

If a failure occurs, the restart time overhead must be taken into account in theglobal execution time. Restart times have been measured and split into its three fun-damental phases explained in Chapter 2: negotiation, le read and eective data re-covery. The negotiation process is measured from the CPPC_Init_configuration()call and until a global agreement about the restart position has been reached, andeach process has identied the state le it must use for restarting. File read time

4.3 Library 135

Table4.19:Runtimeoverhead

caused

bycheckpointing

App.

Runs

xm

inx

max

Min.time

Min.CPPCtime

Max.overhead

NASNPB-MPIv3.1

BT

157138

-2.8233

3.6010

562.87

s.563.43

s.0.6398%

CG

7374

-1.1995

1.2385

211.07

s.210.25

s.0.5868%

EP

213238

0.5300

0.8049

58.71s.

58.32s.

1.3710%

FT

132148

2.0734

3.6943

256.81

s.259.14

s.1.4385%

IS232227

1.9252

2.4136

25.76s.

27.66s.

9.3696%

LU

10689

-2.7466

5.0750

857.30

s.861.99

s.0.5920%

MG

249218

-1.3841

0.9145

66.07s.

66.87s.

1.3841%

SP47

63

1.8426

3.1817

774.12

s.775.48

s.0.4110%

CESG

AFekete

247227

-0.1340

0.1874

51.63s.

52.02s.

0.3630%

CalcuNetw

9287

-3.5301

3.9988

351.71

s.350.23

s.1.1370%

136 Chapter 4. Experimental Results

Table 4.20: Runtime overhead on large-scale applicationsApplication Min. time Overhead Overhead percentage

DBEM 80473.13 s. 256.02 s. 0.31%STEM-II 21622.41 s. 101.14 s. 0.47%

measurements begin when the negotiation ends, and comprise all steps taken untilthe checkpoint data are loaded into memory and made available for the applica-tion to recover them. These include the identication of the writing plugin to beused, le opening and data reading. The recovery begins when the le read ends,and stops when CPPC determines that the restart process has nished, switchingto checkpoint operation mode. This happens when the execution ow reaches thecheckpoint statement where the le was generated in the original execution.

As with the measurements of checkpoint le creation times, 500 experimentswere performed, outliers were discarded from each series, and a 99% condenceinterval was calculated for each of the phases. The results are shown in Table 4.21.Figure 4.26(a) graphically compares the upper limit of the time intervals for eachof the phases. As can be seen, the negotiation is the most costly phase. This is dueto the fact that negotiation times have been measured for checkpoint les includingCRC-32 codes. During the negotiation, application processes have to check theintegrity of the les proposed for restart. If this experiment is performed withstandard les without CRC-32 codes, negotiation times are reduced to the rangeof milliseconds for all applications. If performed on compressed les with CRC-32codes, the time is reduced due to the les being smaller, thus requiring less time forthe integrity checks to be performed.

The times obtained for the le read phase depend basically on the size of theles themselves, being linearly related. Read times would be increased when usingcompressed les because of the need for data decompression. They may also beincreased when using state les generated on dierent architectures, due to thenecessary format conversions. This eect is shown in Figure 4.26(b), which comparesle read times in the Finis Terrae for both Finis Terrae- and Muxia cluster-generatedstate les. This test also serves to demonstrate portability, since restarts take placecorrectly using externally-generated les. Although the test was performed usingthe same statistical approach as above, only the upper limits of the intervals areshown for simplicity. The relationship between both times depends exclusively onthe amount of data that needs to be converted, but the general trend is that theread time is nearly doubled when compared to the original one. The recovery times

4.4 Summary 137

Table 4.21: Restart times (seconds)NEGOTIATION FILE READ RECOVERY

Application xmin xmax xmin xmax xmin xmax

NASNPB-M

PIv3.1 BT 0.2668 0.3027 0.0913 0.0947 0.0507 0.0534

CG 0.2845 0.3177 0.0608 0.0629 0.0551 0.0568EP 0.1714 0.2145 0.0152 0.0162 0.0132 0.0148FT 0.6850 0.7667 0.0911 0.0947 0.1010 0.1033IS 0.5248 0.5330 0.1476 0.1533 0.1783 0.1813LU 0.3675 0.4193 0.0535 0.0553 0.0252 0.0271MG 0.3057 0.3382 0.0500 0.0524 0.0440 0.0461SP 0.2454 0.2580 0.0937 0.0992 0.0903 0.1006

CESGAFekete 0.1775 0.2233 0.0198 0.0210 0.0100 0.0115

CalcuNetw 0.0178 0.0180 0.0111 0.0113 0.0030 0.0030DBEM 3.3010 3.5141 0.6193 0.6498 0.7348 0.7624STEM-II 1.1269 1.1715 0.4711 0.4885 0.5468 0.5792

depend on the amount of data being recovered and the amount of code that mustbe re-executed in order to achieve a complete state recovery.

All these tests show restart times to be very low and fairly negligible in mostsituations. Restart times never exceed a second, except when working with largestate les as is the case with the DBEM and STEM-II applications. Even in thesecases, most of the restart overhead is introduced by the consistency checks on stateles, while the restart protocol itself remains very light and negligible when comparedto the total runtime of these applications.

4.4. Summary

This chapter gives a global overview of the performance obtained by the CPPCframework. A large number of very dierent applications have been used for exper-imental testing, showing adequate automatic processing of them by the frameworkin all cases. Results obtained by the compiler have been thoroughly described, andshown to be adequate in both accuracy of the results and compilation performance,particularly taking into account that the current implementation of the CPPC com-piler has not been thoroughly optimized, neither implemented in a particularly ef-cient environment. After surveying the compilation results, the runtime behaviorof the CPPC library has been addressed. File sizes have been analyzed as one of

138 Chapter 4. Experimental Results

BT CG EP FT IS LU MG SP Fekete CalcuNetw DBEM STEM-II0

1

2

3

4

5

6

Negotiation Read Recovery

time

(se

con

ds)

(a) Upper limit of the condence interval of restart times

BT CG EP FT IS LU MG SP Fekete CalcuNetw DBEM STEM-II0,0

0,2

0,4

0,6

0,8

1,0

1,2

Standard Cross-restart

time

(se

con

ds)

(b) Eect of data conversions in le read times

Figure 4.26: Restart times for test applications

4.4 Summary 139

the fundamental factors in the global performance of the solution, and the benetsobtained by the use of the variable level approach have been addressed. The otherkey factor on checkpointing performance is the writing strategy used. To measurethis impact, state le creation times were measured for all the writing strategiesincluded in the CPPC package. In order to reduce the overhead introduced by thele creation step, CPPC may optionally use a multithreaded dumping algorithm.The global overhead introduced in the test applications when using this approachwas also measured statistically to ensure the accuracy of the results. These haveshown excellent checkpointing overheads, even when the presented numbers were theworst-case ones, and a much better behavior is to be expected in most executions.The overhead of the restart process has also been analyzed and shown to be negligi-ble in reasonable execution scenarios (i.e. with standard failure rates). Cross-restartexperiments have been executed, demonstrating low overheads for data conversionsand the portability of the tool. All runtime tests, with the exception of the runsof large-scale applications, were performed on the Finis Terrae machine hosted bythe Galician Supercomputing Center (CESGA). This supercomputer was ranked as#209 in the June 2008 Top 500 list [1]. It represents a typical machine in a su-percomputing center, and performing the experimental evaluation in it provides amore accurate estimation of what can be expected of the CPPC framework in termsof performance. Also, it demonstrates that the framework may be used in a com-puting infrastructure on which the user has no superuser access or administrationprivileges.

Conclusions and future work

CPPC is a portable and transparent checkpointing infrastructure for long-runningmessage-passing applications. It consists of a runtime library containing routines forcheckpointing support, together with a compiler that automates the use of the li-brary. The most remarkable contributions of this framework are:

The compile-time coordination protocol: consistency issues are moved fromruntime to both compile and restart time. At compile time, checkpoints areplaced in safe points. At restart time, a negotiation between the processesinvolved in the execution of the application decides the safe point from whichto restart. Process synchronization required by traditional coordinated check-pointing approaches is transferred to restart time, thus improving scalabil-ity. Moreover, this solution also enhances eency, since both compiling andrestarting an application are less frequent operations than checkpoint genera-tion.

The reduced state le sizes: CPPC works at variable level. It stores onlythose variables that are needed upon application restart. By restricting theamount of saved data, checkpoint les are made smaller and so checkpointingoverhead decreases. This also improves the performance of network transfers,if necessary.

The portable recovery of the application's state: state recovery is achieved ina portable way by means of both the hierarchical data format used for statedumping, and the selective re-execution of sections of non-portable code. Thisre-execution also provides scope for the checkpointing of applications linkedto external libraries. Portability is a very interesting trait due to the inher-ent heterogeneity of current trends in high performance computing, such asGrid computing. CPPC-G [55] is an ongoing project developing an architec-

141

142 Conclusions and future work

ture based on web services to manage the execution of CPPC fault tolerantapplications on the Grid.

The modular design of the library: the CPPC library is implemented in C++,using a highly modular design that allows for the exible conguration of all itsfunctionalities. It allows for dynamic conguration of le dumping formats, aswell as selection of several capabilities such as le compression, critical whenworking with slow networks or limited amounts of disk space, or integritychecks. This highly modular design also allows for the use of the frameworkin environments where users do not have administrative rights, which is oftenthe case in high performance computers. The use of the MVC pattern allowsthe use of the library from virtually any host programming language, throughthe implementation of thin software layers that adapt the application requeststo the internal interface of the CPPC core.

The fully transparent checkpointing: the CPPC compiler transforms the CPPClibrary into a totally transparent checkpointing tool, through the automationof all required code analyses and transformations. It includes a novel tech-nique for automatic identication of safe points and computation-intensiveloops based on the analysis of code complexity metrics, typical of softwareengineering approaches.

CPPC has been experimentally tested, demonstrating usability, scalability, ef-ciency and portability. It correctly performed application instrumentation androllback-recovery for all the test cases, even using the same set of checkpoint les toperform restart on binary incompatible machines, with dierent C/Fortran compilersand MPI implementations. The experiments were performed on a publicly availablecomputing infrastructure (the Finis Terrae supercomputer hosted by the CESGA),adopting statistical approaches to accurately provide performance estimations.

To our knowledge, CPPC is the only publicly available portable checkpointerfor message-passing applications. CPPC is an open-source project, available athttp://cppc.des.udc.es under GPL license.

This work has spawned the following publications:

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. Contro-lador/preCompilador de Checkpoints Portables. In Actas de las XV Jornadas

de Paralelismo, pp. 253258, Almería (Spain), September 2004.

Conclusions and future work 143

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. On de-signing portable checkpointing tools for large-scale parallel applications. InProceedings of the 2nd International Conference on Computational Science

and Engineering (ICCSE'05), pp. 191203. Istanbul (Turkey), June 2005.

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. PortableCheckpointing of MPI applications. In Proceedings of the 12th Workshop on

Compilers for Parallel Computers (CPC'06), pp. 396410, A Coruña (Spain),January 2006.

G. Rodríguez, M.J. Martín, P. González y J. Touriño. Controller/Precompilerfor Portable Checkpointing. IEICE Transactions on Information and Systems,E89-D(2):408417, February 2006.

G. Rodríguez, M.J. Martín, P. González, J. Touriño y R. Doallo. CPPC: Unaherramienta portable para el checkpointing de aplicaciones paralelas. Boletín

de la red nacional de I+D, RedIRIS, (80):5771, April 2007.

D. Díaz, X.C. Pardo, M.J. Martín, P. González y G. Rodríguez. CPPC-G: Fault-Tolerant Parallel Applications on the Grid. In Proceedings of the

1st Iberian Grid Infrastructure Conference (IBERGRID'07), pp. 230241,Santiago de Compostela (Spain), May 2007.

G. Rodríguez, P. González, M.J. Martín y J. Touriño. Enhancing Fault-Tolerance of Large-Scale MPI Scientic Applications. In Proceedings of the

9th International Conference on Parallel Computing Technologies (PaCT'07),Pereslavl-Zalessky (Russia). Lecture Notes in Computer Science, 4671:153161, September 2007.

D. Díaz, X.C. Pardo, M.J. Martín, P. González y G. Rodríguez. CPPC-G:Fault Tolerant Parallel Applications on the Grid. In 3rd Workshop on LargeScale Computations on Grids (LaSCoG'07), Gda«sk (Poland). Lecture Notes

in Computer Science, 4967:852859, May 2008.

G. Rodríguez, X.C. Pardo, M.J. Martín, P. González, D. Díaz. A FaultTolerance Solution for Sequential and MPI Applications on the Grid. ScalableComputing: Practice and Experience, 9(2):101109, June 2008.

G. Rodríguez, M.J. Martín, P. González, J. Touriño, R. Doallo. CPPC: ACompiler-Assisted Tool for Portable Checkpointing of Message-Passing Appli-cations. In Proceedings of the 1st International Workshop on Scalable Tools for

144 Conclusions and future work

High-End Computing (STHEC'08), held in conjunction with the 22nd ACM In-

ternational Conference on Supercomputing (ICS'08), pp. 112, Kos (Greece),June 2008.

Future work

There are two main improvements that can be made to the CPPC framework. Inthe rst place, the spatial coordination approach used has an important drawback:the need to specify checkpointing frequency in terms of the number of calls to thecheckpointing function, instead of on a temporal basis. We are currently workingon a lightweight, uncoordinated messaging protocol to allow processes to dynami-cally vary the checkpointing frequency based on a user-specied temporal frequency.This protocol involves performing all-to-all non-blocking communications regardingthe speed at which each of the processes progresses through the application code,and adopting the necessary spatial frequency to ensure that the slowest processcheckpoints at the user-specied rate.

The second improvement to be performed is related to the complexity analysisfor checkpoint insertion. The current algorithm implemented in the CPPC compilercan be improved by using more complex metrics, in order to obtain a better approachto the optimal results.

Bibliography

[1] Top 500 supercomputer sites. http://www.top500.org. Last accessed September2008.

[2] Fault tolerance under UNIX. ACM Transactions on Computing, 7(1):124,1989.

[3] N. Aeronautics and S. Administration. The NAS Parallel Benchmarks.http://www.nas.nasa.gov/Software/NPB. Last accessed September 2008.

[4] S. Agarwal, R. Garg, M. Gupta, and J. Moreira. Adaptive incremental check-pointing for massively parallel systems. In Proceedings of the 18th Annual

International Conference on Supercomputing (ICS'04), pages 277286, 2004.

[5] A. Agbaria and R. Friedman. Starsh: Fault-tolerant dynamic MPI programson clusters of workstations. Cluster Computing, 6(3):227236, 2003.

[6] A. Alexandrescu. Modern C++ Design: Generic Programming and Design

Patterns Applied. Addison-Wesley, 1st edition, 2001.

[7] L. Alvisi. Understanding the message logging paradigm for masking process

crashes. PhD thesis, Cornell University, Department of Computer Science,1996.

[8] A. Baratloo, P. Dasgupta, and Z. Kedem. CALYPSO: A novel software systemfor fault-tolerant parallel processing on distributed platforms. In Proceedings

of the 4th IEEE International Symposium on High Performance Distributed

Computing (HPDC-4), pages 122129, 1995.

[9] R. Batchu, J. Neelamegam, C. Zhenqian, M. Beddhu, A. Skjellum, Y. Dandass,and M. Apte. MPI/FTTM: Architecture and taxonomies for fault-tolerant,message-passing middleware for performance-portable parallel computing. In

145

146 BIBLIOGRAPHY

Proceedings of the 1st IEEE International Symposium of Cluster Computing

and the Grid (CCGrid'01), pages 2633, 2001.

[10] A. Beguelin, E. Seligman, and P. Stephan. Application level fault tolerancein heterogeneous networks of workstations. Journal of Parallel and Distributed

Computing, 43(2):147155, 1997.

[11] B. Bhargava and S. Lian. Independent checkpointing and concurrent rollbackfor recovery An optimistic approach. In Proceedings of the 7th Symposium on

Reliable Distributed Systems (SRDS'88), pages 312, 1988.

[12] A. Bouteiller, F. Capello, T. Hérault, G. Krawezik, P. Lemarinier, and F. Mag-niette. MPICH-V2: A fault-tolerant MPI for volatile nodes based on pessimisticsender based message logging. In Proceedings of the 15th ACM/IEEE Confer-

ence on Supercomputing (SC'03), pages 2542, 2003.

[13] G. Bronevetsky, D. Marques, K. Pingali, and P. Stodghill. Automatedapplication-level checkpointing of MPI programs. In Proceedings of the

2003 ACM Symposium on Principles and Practice of Parallel Programming

(PPoPP'03), pages 8494, 2003.

[14] G. Bronevetsky, D. Marques, K. Pingali, and P. Stodghill. C3: A system forautomating application-level checkpointing of MPI programs. In Proceedings

of the 16th International Workshop on Languages and Compilers for Parallel

Computing (LCPC'03), pages 357373, 2003.

[15] G. Bronevetsky, D. Marques, K. Pingali, and P. Stodghill. Collective operationsin application-level fault-tolerant MPI. In Proceedings of the 17th International

Conference on Supercomputing (ICS'03), pages 234243, 2003.

[16] F. Buschmann, R. Meunier, H. Rohnert, P. Sommerlad, and M. Stal. Pattern-Oriented Software Architecture, Volume 1: A System of Patterns. John Wiley& Sons, 1st edition, 1996.

[17] A. Carmona, A. Encinas, and J. Gesto. Estimation of Fekete points. Journalof Computational Physics, 225(2):23542376, 2007.

[18] K. Chandy and L. Lamport. Distributed snapshots: Determining global statesof distributed systems. ACM Transactions on Computer Systems, 3(1):6375,1985.

BIBLIOGRAPHY 147

[19] Y. Chen, J. Plank, and K. Li. CLIP: A checkpointing tool for message-passing

parallel programs, pages 182200. The MIT Press, 2004.

[20] S.-E. Choi and S. Deitz. Compiler support for automatic checkpointing. In Pro-ceedings of the 16th International Symposium on High Performance Computing

Systems and Applications (HPCS'02), pages 213220, 2002.

[21] F. Cristian and F. Jahanian. A timestamp based checkpointing protocol forlong-lived distributed computations. In Proceedings of the 10th Symposium on

Reliable Distributed Systems (SRDS'91), pages 1220, 1991.

[22] E. Elnozahy, L. Alvisi, Y.-M. Wang, and D. Johnson. A survey of rollback-recovery protocols in message-passing systems. ACM Computing Surveys,34(3):375408, 2002.

[23] E. Elnozahy and J. Plank. Checkpointing for Peta-Scale systems: A look intothe future of practical rollback-recovery. IEEE Transactions on Dependable and

Secure Computing, 1(2):97108, 2004.

[24] G. Fagg and J. Dongarra. FT-MPI: Fault tolerant MPI, supporting dynamicapplications in a dynamic world. Lecture Notes in Computer Science, 1908:346353, 2000.

[25] S. Feldman and C. Brown. Igor: A system for program debugging via reversibleexecution. ACM SIGPLAN Notices, 24(1):112123, 1989.

[26] R. Fernandes, K. Pingali, and P. Stodghill. Mobile MPI programs in compu-tational Grids. In Proceedings of the 11th ACM Symposium on Principles and

Practice of Parallel Computing (PPoPP'06), pages 2231, 2006.

[27] N. C. for Supercomputing Applications. HDF-5: File Format Specication.http://hdf.ncsa.uiuc.edu/HDF5/doc/. Last accessed September 2008.

[28] E. Gamma, R. Helm, R. Jonhson, and J. Vlissides. Design Patterns. Elements

of Reusable Object-Oriented Software. Addison-Wesley, 1st edition, 1994.

[29] R. Gioiosa, J. Sancho, S. Jiang, and F. Petrini. Transparent, IncrementalCheckpointing at Kernel Level: a Foundation for Fault Tolerance for Paral-lel Computers. In Proceedings of the 2005 ACM/IEEE Conference on High

Performance Networking and Computing (SC'05), pages 923, 2005.

148 BIBLIOGRAPHY

[30] P. González, T. Pena, and J. Cabaleiro. Dual BEM for crack growth analy-sis on distributed-memory multiprocessors. Advances in Engineering Software,31(12):921927, 2000.

[31] W. Gropp and E. Lusk. Fault Tolerance in Message Passing Interface Programs.International Journal of High Performance Computing Applications, 18(3):363372, 2004.

[32] M. Hayden. The Ensemble system. Technical Report TR98-1662, Cornell Uni-versity, Department of Computer Sciences, 1998.

[33] J. Hélary, A. Mostefaoui, and M. Raynal. Virtual precedence in asynchronoussystems: concepts and applications. In Proceedings of the 11th Workshop on

Distributed Algorithms (WDAG'97), pages 170184, 1997.

[34] D. Johnson and W. Zwaenepoel. Sender-based message logging. In Proceed-

ings of the 17th Annual International Symposium on Fault-Tolerant Computing

(FTCS'87), pages 1419, 1987.

[35] R. Koo and S. Toueg. Checkpointing and rollback-recovery for distributedsystems. IEEE Transactions on Software Engineering, 13(1):2331, 1987.

[36] C. Landau. The checkpoint mechanism in KeyKOS. In Proceedings of the

2nd International Workshop on Object Orientation on Operating Systems (I-

WOOOS'92), pages 8691, 1992.

[37] S.-I. Lee, T. Johnson, and R. Eigenmann. Cetus an extensible compilerinfrastructure for source-to-source transformation. In Proceedings of the 16th

International Workshop on Languages and Compilers for Parallel Computing

(LCPC'03), pages 539553, 2003.

[38] C.-C. Li, E. Stewart, and W. Fuchs. Compiler-assisted full checkpointing. Soft-ware: Practice and Experience, 24(10):871886, 1994.

[39] C.-C. Li, E. Stewart, and W. Fuchs. Compiler-assisted full checkpointing. Soft-ware: Practice and Experience, 24(10):871886, 1994.

[40] M. Litzkow, M. Livny, and M. Mutka. Condor A hunter of idle workstations.In Proceedings of the 8th International Conference on Distributed Computing

Systems (ICDCS'88), pages 104111, 1988.

BIBLIOGRAPHY 149

[41] M. Litzkow, T. Tannenbaum, J. Basney, and M. Livny. Checkpoint and migra-tion of UNIX processes in the Condor distributed processing system. TechnicalReport #1346, University of Wisconsin-Madison, 1997.

[42] M. Martín, D. Singh, J. Mouriño, F. Rivera, R. Doallo, and J. Bruguera. Highperformance air pollution modeling for a power plant environment. Parallel

computing, 29(1112):17631790, 2003.

[43] J. Mouriño, E. Estrada, and A. Gómez. CalcuNetw: Calculate measurementsin complex networks. Technical Report 2005-003, Galician SupercomputingCenter (CESGA), 2005.

[44] R. Netzer and J. Xu. Necessary and sucient conditions for consistent globalsnapshots. IEEE Transactions on Parallel and Distributed Systems, 6(2):165169, 1995.

[45] J. Ousterhout, A. Cherenson, F. Douglis, M. Nelson, and B. Welch. The Spritenetwork operating system. IEEE Computer, 21(2):2336, 1988.

[46] T. Parr. Antlr parser generator. http://www.antlr.org. Last accessed September2008.

[47] J. Plank. Ecient checkpointing on MIMD architectures. PhD thesis, PrincetonUniversity, Department of Computer Science, 1993.

[48] J. Plank, M. Beck, and G. Kingsley. Compiler-assisted memory exclusion forfast checkpointing. IEEE Technical Committee on Operating Systems and Ap-

plication Environments, 7(4):1014, 1995.

[49] J. Plank, M. Beck, G. Kingsley, and K. Li. Libckpt: Transparent checkpointingunder Unix. In Usenix Winter Technical Conference, pages 213223, 1995.

[50] J. Plank, J. Xu, and R. Netzer. Compressed dierences: An algorithm for fastincremental checkpointing. Technical Report CS-95-302, University of Ten-nessee, 1995.

[51] B. Ramkumar and V. Strumpen. Portable checkpointing for heterogeneousarchitectures. In Proceedings of the 27th International Symposium on Fault-

Tolerant Computing (FTCS'97), pages 5867, 1997.

[52] B. Randell. System structure for software fault tolerance. IEEE Transactions

on Software Engineering, 1(2):221232, 1975.

150 BIBLIOGRAPHY

[53] S. Rao, L. Alvisi, and H. Vin. Egida: an extensible toolkit for low-overheadfault-tolerance. In Proceedings of the 29th Annual International Symposium on

Fault-Tolerant Computing (FTCS'99), pages 4855, 1999.

[54] G. Rodríguez, M. Martín, P. González, and J. Touriño. Controller/Precompilerfor Portable Checkpointing. IEICE Transactions on Information and Systems,E89-D(2):408417, 2006.

[55] G. Rodríguez, X. Pardo, M. Martín, P. González, and D. Díaz. A fault tolerancesolution for sequential and MPI applications on the Grid. Scalable Computing:Practice and Experience, 9(2):101109, 2008.

[56] M. Russinovich and Z. Segall. Fault-tolerance for o-the-shelf applicationsand hardware. In Proceedings of the 25th International Symposium on Fault-

Tolerant Computing (FTCS'95), pages 6771, 1995.

[57] M. Sezgin and B. Sankur. Survey over image thresholding techniques andquantitative performance evaluation. Journal of Electronic Imaging, 13(1):146165, 2004.

[58] D. Shires, L. Pollock, and S. Sprenkle. Program ow graph construction forstatic analysis of MPI programs. In Proceedings of the International Con-

ference on Parallel and Distributed Processing Techniques and Applications

(PDPTA'99), pages 18471853, 1999.

[59] S. Siegel and G. Avrunin. Modeling wilcard-free MPI programs for verica-tion. In Proceedings of the 10th ACM Symposium on Principles and Practice

of Parallel Computing (PPoPP'05), pages 95106, 2005.

[60] S. Smale. Mathematical problems for the next century. Mathematical Intelli-

gencer, 20(2):715, 1998.

[61] G. Stellner. Cocheck: Checkpointing and process migration for MPI. In Pro-

ceedings of the 10th International Parallel Processing Symposium (IPPS'96),pages 526531, 1996.

[62] R. Strom and S. Yemini. Optimistic recovery in distributed systems. ACM

Transactions on Computer Systems, 3(3):204226, 1985.

[63] B. Stroustrup. The C++ Programming Language. Addison-Wesley, 3rd edition,2000.

BIBLIOGRAPHY 151

[64] V. Strumpen. Portable and fault-tolerant software systems. IEEE Micro,18(5):2232, 1998.

[65] V. Sunderam. PVM: A framework for parallel distributed computing. Concur-rency, Practice and Experience, 2(4):315340, 1990.

[66] Y. Tamir and C. Sequin. Error recovery in multicomputers using global check-points. In Proceedings of the International Conference on Parallel Processing

(ICPP'84), pages 3241, 1984.

[67] S. Toueg and Ö. Babaoglu. On the optimum checkpoint selection problem.SIAM Journal on Computing, 13(3):630649, 1984.

[68] J. Tukey. Exploratory data analysis. Addison-Wesley, 1977.

[69] N. Vaidya. Impact of checkpoint latency on overhead ratio of a checkpointingscheme. IEEE Transactions on Computers, 46(8):942947, 1997.

[70] Y.-M. Wang. Checkpoint space reclamation for uncoordinated checkpointing in

message-passing systems. PhD thesis, University of Illinois, Department ofComputer Science, 1993.

[71] N. Woo, H. Jung, H. Yeom, T. Park, and H. Park. MPICH-GF: Transparentcheckpointing and rollback-recovery for Grid-enabled MPI processes. IEICE

Transactions on Information and Systems, E87-D(7):18201828, 2004.

[72] Xerces C++ Parser. http://xml.apache.org/xerces-c/. Last accessed September2008.

[73] G. Zack, W. Rogers, and S. Latt. Automatic measurement of sister chromatidexchange frequency. Journal of Histochemistry & Cytochemistry, 25(7):741753,1977.

Compiler-assisted checkpointing of message-passing …gac.udc.es/tesis/GabrielRodriguezAlvarez.pdf · 2011-12-13 · Ejemplos de estas herramientas son Calypso [8] y las extensiones - [PDF Document] (2024)

References

Top Articles
Latest Posts
Recommended Articles
Article information

Author: Ouida Strosin DO

Last Updated:

Views: 5337

Rating: 4.6 / 5 (56 voted)

Reviews: 87% of readers found this page helpful

Author information

Name: Ouida Strosin DO

Birthday: 1995-04-27

Address: Suite 927 930 Kilback Radial, Candidaville, TN 87795

Phone: +8561498978366

Job: Legacy Manufacturing Specialist

Hobby: Singing, Mountain biking, Water sports, Water sports, Taxidermy, Polo, Pet

Introduction: My name is Ouida Strosin DO, I am a precious, combative, spotless, modern, spotless, beautiful, precious person who loves writing and wants to share my knowledge and understanding with you.