Vers une meilleure utilisation des ressources d’E/S en HPC
Francieli Boito · HAL (Le Centre pour la Communication Scientifique Directe) · 2025
In high-performance computing (HPC) systems, input/output (I/O) resources are not traditionally arbitrated in the same way as compute resources (nodes, CPUs, etc). However, I/O resources havean important impact on performance, and this impact depends on application characteristics. Moreover, these resources are potentially shared by jobs, which may then suffer with interference from others. In this document, I summarize my research efforts from the past decade (since my thesis) towardsimproving I/O resource management. I present this work classified into three parts: the allocation,scheduling, and placement problems. For each part, I discuss proposed techniques and perspectivesfor future work. Moreover, transversal to the whole discussion is the problem of characterizing applications to provide information for resource management techniques. Therefore, throughout the text,the strategies proposed for this characterization are detailed and discussed.