Reliability Analysis in HPC clusters

Resource failures and down times have become a growing concern for large-scale computational platforms, as they tend to have an adverse affect on the performance of the computation system,. Reliabilityaware resource allocation and checkpointing algorithms have been investigated to minimize the performance loss due to unexpected failures. The effectiveness of a reliability-aware policy relies on the accuracy of reliability prediction. The reliability of a group of nodes is evaluated as a combination of individual node information under an assumption that each node reliability is independent. In this paper, we describe the reliability analysis based on time between failures for a system/group of nodes. Various reliability models are compared for different cases in order to find an optimal reliability model. The reliability models and analysis techniques were evaluated based on actual failure data of 512 nodes during their four year operational period. 1 1