Dynamic Resource Management (DRM) facilitates the real-time adjustment of resources allocated to a job during its execution. Recently, DRM has gained interest due to its potential to contribute to sustainable HPC by enhancing energy efficiency and increasing throughput for HPC system providers and users. Implementing DRM requires comprehensive changes and coordination across the entire HPC software stack. We explore a layered approach for DRM highlighting ongoing work on various solutions for DRM across the software stack. This includes the OAR resource manager, the Dynamic Processes with PSets (DPP) approach, the Dynamic Management of Resources (DMR) framework, and adaptive SWEET/LibPFASST. We showcase how these approaches create a cohesive strategy for DRM across the whole HPC software stack, demonstrating our collaborative effort for an integrated solution.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A Layered Approach for Dynamic Resource Management in HPC

  • Hans-Joachim Bungartz,
  • Pierre-François Dutot,
  • Jan Fecht,
  • Keerthi Gaddameedi,
  • Dominik Huber,
  • Sergio Iserte,
  • Michael Minion,
  • Tobias Neckel,
  • Antonio Peña,
  • Olivier Richard,
  • Martin Schreiber,
  • Martin Schulz,
  • Valentina Schüller

摘要

Dynamic Resource Management (DRM) facilitates the real-time adjustment of resources allocated to a job during its execution. Recently, DRM has gained interest due to its potential to contribute to sustainable HPC by enhancing energy efficiency and increasing throughput for HPC system providers and users. Implementing DRM requires comprehensive changes and coordination across the entire HPC software stack. We explore a layered approach for DRM highlighting ongoing work on various solutions for DRM across the software stack. This includes the OAR resource manager, the Dynamic Processes with PSets (DPP) approach, the Dynamic Management of Resources (DMR) framework, and adaptive SWEET/LibPFASST. We showcase how these approaches create a cohesive strategy for DRM across the whole HPC software stack, demonstrating our collaborative effort for an integrated solution.