HydEE: Failure Containment without Event Logging for Large Scale Send-Deterministic MPI Applications

Guermouche, Amina; Ropars, Thomas; Snir, Marc; Cappello, Franck

doi:10.1109/ipdps.2012.111

Cited by 34 publications

(48 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This induces an overhead, which we express as a slowdown of the execution rate: instead of executing one work-unit per second, the application executes only λ work-units, where 0 < λ < 1. Typical values for λ are said to be λ ≈ 0.98, meaning that the overhead due to payload messages is only a small percentage [36,13]. On the contrary, message logging has a positive effect on re-execution after a failure, because intergroup messages are stored in memory and directly accessible after the recovery.…”

Section: Refining the Modelmentioning

confidence: 99%

See 1 more Smart Citation

Fault-Tolerance Techniques for High-Performance Computing

2015

Computer Communications and Networks

View full text Add to dashboard Cite

This report provides an introduction to resilience methods. The emphasis is on checkpointing, the de-facto standard technique for resilience in High Performance Computing. We present the main two protocols, namely coordinated checkpointing and hierarchical checkpointing. Then we introduce performance models and use them to assess the performance of theses protocols. We cover the Young/Daly formula for the optimal period and much more! Next we explain how the efficiency of checkpointing can be improved via fault prediction or replication. Then we move to application-specific methods, such as ABFT. We conclude the report by discussing techniques to cope with silent errors (or silent data corruption).This report is a slightly modified version of the first chapter of the monograph Fault tolerance techniques for high-performance computing edited by Thomas Herault and Yves Robert, and to be published by Springer Verlag.

show abstract

Section: Refining the Modelmentioning

confidence: 99%

“…Typically, these groups are composed to take advantage of the application communication pattern [36,32]. For instance, if the application executes on a 2D-grid of processors, a natural way to create processor groups is to have one group per row (or column) of the grid.…”

Section: Case Studiesmentioning

confidence: 99%

Fault-Tolerance Techniques for High-Performance Computing

2015

Computer Communications and Networks

View full text Add to dashboard Cite

show abstract

“…Several novel fault tolerance protocols overcome this limitation by reducing significantly the number of messages to log. They fall into the class of hierarchical fault tolerance protocols, forming clusters of processes and using coordinated checkpointing inside clusters and and message logging between clusters [61,77,84]. Such protocols need to manage causal dependencies between processes in order to ensure correct recovery.…”

Section: Toward Exascale Resilience: 2014 Updatementioning

confidence: 99%

Toward Exascale Resilience: 2014 update

Cappello

Geist

Gropp

et al. 2014

JSFI

101

View full text Add to dashboard Cite

Resilience is a major roadblock for HPC executions on future exascale systems. These systems will typically gather millions of CPU cores running up to a billion threads. Projections from current large systems and technology evolution predict errors will happen in exascale systems many times per day. These errors will propagate and generate various kinds of malfunctions, from simple process crashes to result corruptions.The past five years have seen extraordinary technical progress in many domains related to exascale resilience. Several technical options, initially considered inapplicable or unrealistic in the HPC context, have demonstrated surprising successes. Despite this progress, the exascale resilience problem is not solved, and the community is still facing the difficult challenge of ensuring that exascale applications complete and generate correct results while running on unstable systems. Since 2009, many workshops, studies, and reports have improved the definition of the resilience problem and provided refined recommendations. Some projections made during the previous decades and some priorities established from these projections need to be revised. This paper surveys what the community has learned in the past five years and summarizes the research problems still considered critical by the HPC community.

show abstract

“…Section 0.7 is mainly based on [12] which contains references to many studies of checkpointing policies. For non-coordinated checkpointing protocols, [28] by Guermouche et al is a good starting point.…”

Section: Further Informationmentioning

confidence: 99%

Scheduling for Large-Scale Systems

Vivien¹

2014

Computing Handbook, Third Edition

View full text Add to dashboard Cite

In this chapter, scheduling is defined as the activity that consists in mapping a task graph onto a target platform. The task graph represents the application, nodes denote computational tasks, and edges model precedence constraints between tasks. For each task, an assignment (choose the processor that will execute the task) and a schedule (decide when to start the execution) are determined. The goal is to obtain an efficient execution of the application, which translates into optimizing some objective function.The traditional objective function in the scheduling literature is the minimization of the total execution time, or makespan. Section 0.2 provides a quick background on task graph scheduling, mostly to recall some definitions, notations, and well-known results.In contrast to pure makespan minimization, this chapter aims at discussing scheduling problems that arise at large scale, and involve one or several different optimization criteria, that depart from, or complement, the classical makespan objective. Section 0.3 assesses the importance of key ingredients of modern scheduling techniques: (i) multi-criteria optimization; (ii) memory management; and (iii) reliability. These concepts are illustrated in the following sections, which cover four case studies. Section 0.4 is devoted to illustrate multi-criteria optimization techniques, that mix throughput, energy and reliability. Sections 0.5 and 0.6 illustrate different memory management techniques to schedule large task graphs. Finally, Section 0.7 discusses the impact of checkpointing techniques to schedule parallel jobs at very large-scale. We conclude this chapter by stating some research directions in Section 0.8, by defining terms in Section 0.9, and by pointing to additional sources of information in Section 0.10. Background on Scheduling 0.2.1 Makespan MinimizationTraditional scheduling assumes that the target platform is a set of p identical processors, and that no communication cost is paid. In that context, a task graph is a directed acyclic vertex-weighted graph G = (V, E, w), where the set V of vertices represents the tasks, the set E of edges represents precedence constraints between tasks (e = (u, v) ∈ E if and only if u ≺ v). and the weight function w : V −→ N * gives the weight (or duration) of each task. Task weights are assumed to be positive integers. A schedule σ of a task graph is a function that assigns a start time to each task: σ : V −→ N * such that σ(u) + w(u) ≤ σ(v) whenever e = (u, v) ∈ E. In other words, a schedule preserves the dependence constraints induced by the precedence relation ≺ and embodied by the edges of the dependence graph; if u ≺ v, then the execution of u begins at time σ(u) and requires w(u) units of time, and the execution of v at time σ(v) must start after the end 1

show abstract

HydEE: Failure Containment without Event Logging for Large Scale Send-Deterministic MPI Applications

Cited by 34 publications

References 26 publications

Fault-Tolerance Techniques for High-Performance Computing

Fault-Tolerance Techniques for High-Performance Computing

Toward Exascale Resilience: 2014 update

Scheduling for Large-Scale Systems

Contact Info

Product

Resources

About