Dynamic Resources Management for Exascale in ADMIRE Framework

Due to new workloads arising with workflows and AI applications, dynamic resource allocation can improve productivity across all system and application levels by adapting the applications' configurations to the system's resources. However, HPC system software is not suited nowadays to provide dynamic resource management (DRM) for computing or I/O resources in runtime. This paper presents the ADMIRE framework and its mechanisms for dynamic resource management. ADMIRE's efforts enable an integrated stack in which storage, compute, and orchestration layers were co-designed and validated across a shared, heterogeneous testbed.