Algorithm-based fault tolerance applied to high performance computing

Bosilca, George; Delmas, Rémi; Dongarra, Jack; Langou, Julien

doi:10.1016/j.jpdc.2008.12.002

Cited by 186 publications

(147 citation statements)

References 23 publications

Supporting

Mentioning

146

Contrasting

Unclassified

Order By: Relevance

“…With checkpointing, several application-specific detectors can be used to avoid replication and increase performance in failure-free executions. Two well-known examples are AlgorithmBased Fault Tolerance (ABFT) [38,39,40] and silent error detectors based on domain-specific data analytics [41,42,43].…”

Section: Soft and Silent Errorsmentioning

confidence: 99%

Checkpointing Workflows for Fail-Stop Errors

Han¹,

Canon²,

Casanova³

et al. 2017

2017 IEEE International Conference on Cluster Computing (CLUSTER)

View full text Add to dashboard Cite

Abstract:We consider the problem of orchestrating the execution of workflow applications structured as Directed Acyclic Graphs (DAGs) on parallel computing platforms that are subject to fail-stop failures. The objective is to minimize expected overall execution time, or makespan. A solution to this problem consists of a schedule of the workflow tasks on the available processors and of a decision of which application data to checkpoint to stable storage, so as to mitigate the impact of processor failures. For general DAGs this problem is hopelessly intractable. In fact, given a solution, computing its expected makespan is still a difficult problem. To address this challenge, we consider a restricted class of graphs, Minimal Series-Parallel Graphs (M-SPGs). It turns out that many real-world workflow applications are naturally structured as M-SPGs. For this class of graphs, we propose a recursive list-scheduling algorithm that exploits the M-SPG structure to assign sub-graphs to individual processors, and uses dynamic programming to decide which tasks in these sub-gaphs should be checkpointed. Furthermore, it is possible to efficiently compute the expected makespan for the solution produced by this algorithm, using a first-order approximation of task weights and existing evaluation algorithms for 2-state probabilistic DAGs. We assess the performance of our algorithm for production workflow configurations, comparing it to (i) an approach in which all application data is checkpointed, which corresponds to the standard way in which most production workflows are executed today; and (ii) an approach in which no application data is checkpointed. Our results demonstrate that our algorithm strikes a good compromise between these two approaches, leading to lower checkpointing overhead than the former and to better resilience to failure than the latter. To the best of our knowledge, this is the first scheduling/checkpointing algorithm for workflow applications with fail-stop failures that considers workflow structures more general than mere linear chains of tasks.Key-words: workflow, checkpoint, fail-stop error, resilience. Stratégies de checkpoint pour les workflows en présence d'erreurs fatalesRésumé : Ce rapport considère l'ordonnancement de workflows (applications structurées en forme de graphes de tâches acycliques, ou DAGs) sur des plates-formes parallèlesà grandé echelle, soumisesà des erreurs fatales. L'objectif est de minimiser l'espérance du temps total d'exécution, ou makespan. Une solutionà ce problème comprend l'allocation ordonnée des tâches aux processeurs, et les décisions de checkpoint: quelles tâches sont suivies d'un checkpoint? Même pour une solution donnée, le calcul du makespan reste difficile. Nous nous restreignonsà une classe de DAGs particuliers, les graphes séries-parallèles minimaux, ou MSPGs. De nombreux workflows issus des applications ont pour graphe un M-SPG. Pour de tels graphes, nous proposons un algorithme qui utilise la structure récursive du M-SPG pour allouer des sous-graphesà chaque pro...

show abstract

Section: Soft and Silent Errorsmentioning

confidence: 99%

Checkpointing Workflows for Fail-Stop Errors

Han¹,

Canon²,

Casanova³

et al. 2017

2017 IEEE International Conference on Cluster Computing (CLUSTER)

View full text Add to dashboard Cite

show abstract

“…Strong scalability involves doubling the number of nodes but maintaining the constant configuration size for the chosen scientific code, while weak scalability occurs when the number of nodes is doubled and the problem size is also concurrently increased (Bosilca, Delmas, Dongarra, & Langou, 2009;Varma, Wang, Mueller, Engelmann, & Scott, i2006). Using DL_POLY_2.18, the strong scalability of the model was studied in two different ways, namely, by using a small and large system (Tang, 2007).…”

Section: Design Of Computational Experimentsmentioning

confidence: 99%

“…In the small system, DL_POLY_2 was employed for a simulation comprising 8640 atoms (that is Sodium (Na) = 960, Potassium (K) = 960, Silicon (Si) = 1920, Oxygen (O) = 4800) (Bosilca et al, 2009). The chemical system of atoms were simulated in a cubic box of size 48.358 angstroms for each of its X, Y and Z axes.…”

Section: Design Of Computational Experimentsmentioning

confidence: 99%

Scalability of DL_POLY on High Performance Computing Platform

Mabakane

Moeketsi

Lopis

2017

SACJ

View full text Add to dashboard Cite

This paper presents a case study on the scalability of several versions of the molecular dynamics code (DL_POLY) performed on South Africa‘s Centre for High Performance Computing e1350 IBM Linux cluster, Sun system and Lengau supercomputers. Within this study different problem sizes were designed and the same chosen systems were employed in order to test the performance of DL_POLY using weak and strong scalability. It was found that the speed-up results for the small systems were better than large systems on both Ethernet and Infiniband network. However, simulations of large systems in DL_POLY performed well using Infiniband network on Lengau cluster as compared to e1350 and Sun supercomputer.

show abstract

“…Indeed, application-specific information enables adhoc solutions, which dramatically decrease the cost of error detection. Algorithmbased fault tolerance (ABFT) [29,30,31] is a well-known technique, which uses checksums to detect up to a certain number of errors in linear algebra kernels. Unfortunately, ABFT can only protect datasets in linear algebra kernels, and it must be implemented for each different kernel, which incurs a large amount of work for large HPC applications.…”

Section: Related Workmentioning

confidence: 99%

Resilient N-Body Tree Computations with Algorithm-Based Focused Recovery: Model and Performance Analysis

Cavelan

Fang

Chien

et al. 2017

Lecture Notes in Computer Science

View full text Add to dashboard Cite

This paper presents a model and performance study for Algorithm-Based Focused Recovery (ABFR) applied to N-body computations, subject to latent errors. We make a detailed comparison with the classical Checkpoint/Restart (CR) approach. While the model applies to general frameworks, the performance study is limited to perfect binary trees, due to the inherent difficulty of the analysis. With ABFR, the crucial parameter is the detection interval, which bounds the error latency. We show that the detection interval has a dramatic impact on the overhead, and that optimally choosing its value leads to significant gains over the CR approach.

show abstract

Algorithm-based fault tolerance applied to high performance computing

Cited by 186 publications

References 23 publications

Checkpointing Workflows for Fail-Stop Errors

Checkpointing Workflows for Fail-Stop Errors

Scalability of DL_POLY on High Performance Computing Platform

Resilient N-Body Tree Computations with Algorithm-Based Focused Recovery: Model and Performance Analysis

Contact Info

Product

Resources

About