Optimizing HPC Fault-Tolerant Environment: An Analytical Approach

Jin, Hui; Chen, Yong; Zhu, Huaiyu; Sun, Xian-He

doi:10.1109/icpp.2010.80

Cited by 37 publications

(30 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, it worth noting that such result will be impossible to use in the context of HPC due the huge number of processors involved in the computation given that the computational complexity of the algorithm provided by Plank et al is exponential with respect to the number of processors. In [19], Jin et al provided an analytical model to express the expected completion time in presence of failures with respect to the number of the involved processors in the computation and the amount of workload to execute. Using the queuing theory and by applying Taylor expansion to the expression of the expected completion time of the application with respect to the involved processors in the computation they provided a first order approximation to the optimal checkpoint interval considering that the checkpoint cost is constant.…”

Section: Related Workmentioning

confidence: 99%

Complexity Analysis of Checkpoint Scheduling with Variable Costs

Bouguerra

Trystram²,

Wagner

2013

IEEE Trans. Comput.

View full text Add to dashboard Cite

International audienceThe parallel computing platforms available today are increasingly larger and thus, more and more subject to failures. Consequently it is necessary to develop efficient strategies providing safe and reliable completion for HPC parallel applications. Checkpointing is one of the most popular and efficient technique for developing fault-tolerant applications on such context. However, checkpoint operations are costly in terms of time, computation and network communication. This will certainly affect the global performance of the application. In this work, we propose a performance model that expresses formally the checkpoint scheduling problem. This model exhibits the tradeoff between the impact of the checkpoints operations and the lost computation due to failures. Based on this model we study the computational complexity of the problem of scheduling checkpoints with variable costs for general failure distributions. More precisely, we provide a new computational complexity analysis which explicits in depth the relations between the probabilistic failure model, the checkpoint cost and the computational model. In particular, we prove that the checkpoint scheduling problem is NP-hard even in the simple case of uniform failure distribution. We also present a dynamic programming scheme for determining the optimal checkpointing times in all the variants of the problem

show abstract

Section: Related Workmentioning

confidence: 99%

Complexity Analysis of Checkpoint Scheduling with Variable Costs

Bouguerra

Trystram²,

Wagner

2013

IEEE Trans. Comput.

View full text Add to dashboard Cite

show abstract

“…Balaprakash et al [31] address the question of optimizing the checkpoint intervals while also considering the energy consumption of a multilevel scheme. As with the models for the coordinated schemes [26], [27] and [28], these hierarchical models cannot be used for non-hierarchical unified uncoordinated and coordinated checkpointing systems such as combined task-level and system-wide checkpointing. The unified model proposed by Bosilca et al [32] models a range of checkpointing systems from fully coordinated scheme on one extreme to partially coordinated hierarchical schemes on the other extreme.…”

Section: Related Workmentioning

confidence: 97%

“…Young and Daly study the sequential jobs [12] [13]. For parallel jobs, studies such as [26], [27] and [28] model the coordinated checkpointing protocols.…”

Section: Related Workmentioning

confidence: 99%

Marriage Between Coordinated and Uncoordinated Checkpointing for the Exascale Era

Subaşi

Zyulkyarov

Ünsal

et al. 2015

2015 IEEE 17th International Conference on High Performance Computing and Communications, 2015 IEEE 7th International Symposium

View full text Add to dashboard Cite

The state-of-the-art checkpointing techniques are projected to be prohibitively expensive in the Exascale era. These techniques are most often holistic in nature which prevents them to leverage programming model and paradigm specific advantages so as to be viable for the Exascale era. In this work, we present a unified non-hierarchical model to combine uncoordinated checkpointing with coordinated system-wide checkpointing to capitalize on programming model specific advantages. We develop closed-form formulas for performance improvement and the optimal checkpoint interval of the unified model in our analytical assessment. As an instantiation of our model, we propose to unify task-level checkpointing with a system-wide checkpointing scheme for task-parallel HPC applications. This instantiation has three distinct advantages: first it reduces performance overheads by decreasing the frequency of checkpoints in the unified system; second it features fast failure recovery by using in-memory tasklocal checkpoints instead of on-disk global checkpoints; and third it does not compromise from the high failure coverage typical of system-wide checkpointing.

show abstract

“…With the data distribution, I/O interconnect usage, and access pattern of applications provided by the DRA component, the DDC component coordinates the I/O accesses to manage the substantial amount of concurrency and to mitigate the contention issue on exascale systems. The DDC component orchestrates the I/O requests in both independent I/O and collective I/O operations, which have been observed with serious contention issue at a large scale [CSTS10,JiCS10]. The dynamic coordination for both independent and collective I/O is discussed in the following subsections.…”

Section: Dynamically Coordinated I/o Architecturementioning

confidence: 99%

Towards scalable I/O architecture for exascale systems

Chen

2011

Proceedings of the 2011 ACM International Workshop on Many Task Computing on Grids and Supercomputers

Self Cite

View full text Add to dashboard Cite

High performance computing (HPC) has crossed the Petaflop mark and is reaching the Exaflop range quickly. The exascale system is projected to have millions of nodes, with thousands of cores for each node. At such an extreme scale, the substantial amount of concurrency can cause a critical contention issue for I/O system. The contention can destroy the request locality, increase the access latency, and waste the precious I/O interconnection bandwidth. This study proposes a dynamically coordinated I/O architecture for exascale systems. The fundamental idea is to coordinate I/O accesses according to access pattern, network topology, interconnection condition, and data distribution on storage devices to reduce the contention and regain the locality. The preliminary results have shown the promise of a dynamically coordinated I/O architecture. It has a potential to manage the substantial amount of I/O concurrency and provides a scalable I/O architecture for exascale systems.

show abstract

Optimizing HPC Fault-Tolerant Environment: An Analytical Approach

Cited by 37 publications

References 22 publications

Complexity Analysis of Checkpoint Scheduling with Variable Costs

Complexity Analysis of Checkpoint Scheduling with Variable Costs

Marriage Between Coordinated and Uncoordinated Checkpointing for the Exascale Era

Towards scalable I/O architecture for exascale systems

Contact Info

Product

Resources

About