Framework and Modular Infrastructure for Automation of Architectural Adaptation and Performance Optimization for HPC Systems

摘要: High performance systems have complex, diverse and rapidly evolving architectures. The span of applications, workloads, resource use patterns is diversifying. Adapting applications for efficient execution on this spectrum environments effort intensive. There are many optimization tools which implement some or several aspects the full task but almost none comprehensive across architectures, environments, workloads. This paper presents, illustrates, applies a modular infrastructure enables composition multiple open-source analyses into set workflows implementing end-to-end HPC architectures types parallel environments. It gives results from an implementation Stampede system at Texas Advanced Computing Center where user can submit application using only single command line get back least, partially optimized program without manual modification two different chips. Currently, subset possible optimizations completely automated growing. Case studies workflow presented. implementations currently available download as PerfExpert tool version 4.0 supports both Sandy Bridge Intel Phi