Modélisation d'allocation de ressources de stockage hétérogènes sur plateformes HPC

International audienceIn recent years, and despite remarkable progress in computing and network performance, HPC platforms have struggled to maintain satisfactory I/O throughput. Various solutions have been proposed to mitigate the contention and variability experienced by more and more concurrent applications, particularly on heavily shared parallel file systems. As a result, many large scale platforms now offer complex hierarchies of storage resources using various architectures (node-local, burst buffers, network-attached storage and so on) based on diversified hardware technologies such as persistent memories or flash for instance. In that context, we propose to study how to efficiently allocate these heterogeneous storage resources for scientific applications and workflows. For that purpose, we developed StorAlloc, a modular and extensible simulator of a storage-aware job-scheduler. We ran a large set of experiments with StorAlloc to investigate storage system designs and resource scheduling algorithms, through the analysis of multiple storagerelated metrics. For example, we have been able to determine an estimate of the size of a burst buffers partition sufficient to capture the intensive I/O of a top-tier supercomputer