SHAMan

Abstract

Like most modern computer systems, High Performance Computing (HPC) machines integrate many highly configurable hardware devices and software components. Finding their optimal parametrization is a complex task, as the size of the parametric space and the non-linear behavior of HPC systems make hand tuning, theoretical modeling or exhaustive sampling unsuitable for most cases. Auto-tuning methods relying on black-box optimization have emerged as a promising solution for finding systems' best parametrization without making any assumption on their behaviors. In this paper, we present the architecture of an auto-tuning framework, called Smart HPC Application MANager (SHAMan), that integrates black-box optimization heuristics to find the optimal parametrization of an Input/Output (I/O) accelerator for a HPC application. We describe the conceptual and technical architecture of the framework and its native support for HPC clusters' ecosystem. We detail in depth the stand-alone optimization engine and its integration as a service provided by a Web application. We deployed and tested the framework by tuning an I/O accelerator developed by the Atos company on a HPC cluster running in production. The tuner's performance is evaluated by optimizing 90 different I/O oriented applications. We show a median improvement of 29% in speed-up compared to the default parametrization and this improvement goes up to 98% for a certain class of applications.