Auto-tuning of IO accelerators using black-box optimization
Sophie Robert, Soraya Zertal, Gaël Goret · 2019
High Performance Computing (HPC) applications' performance and behavior rely on software and hardWare environments Which are often highly configurable. Finding their optimal parametrization is a very complex task. The size of the parametric space and the non-linear relationship between the parameters and the delivered performance make hand-tuning, theoretical modeling or exhaustive sampling unsuitable for most cases. In this paper, We propose an auto-tuning loop that uses black-box optimization to Find the optimal parametrization of IO accelerators for a given HPC application in a limited number of iterations, Without making any assumption on the performance function. After a literature review of the selected methods for tuning the accelerators, We describe their implementation and experimentation in our HPC context using two IO accelerators developed by Atos. We also define several metrics to evaluate the quality of our optimization, as our criteria of success go further than finding the optimal parameters. The obtained results show that this framework successfully improves the execution time of two applications used conjointly With a pure software accelerator and a mixed hardWare-software one. We indeed observe possible time gains of respectively 38% and 20% for each accelerator compared to launching the same application accelerated With the default parameters.