Auto-tuning methodology for configuration and application parameters of hybrid CPU + GPU parallel systems based on expert knowledge

Abstract

Auto-tuning of configuration and application parameters allows to achieve significant performance gains in many contemporary compute-intensive applications. Feasible search spaces of parameters tend to become too big to allow for exhaustive search in the auto-tuning process. Expert knowledge about the utilized computing systems becomes useful to prune the search space and new methodologies are needed in the face of emerging heterogeneous computing architectures. In this paper we propose an auto-tuning methodology for hybrid CPU/GPU applications that takes into account previous execution experiences, along with an automated tool for iterative testing of chosen combinations of configuration, as well as application-related parameters. Experimental results, based on a parallel similarity search application executed on three different CPU + GPU parallel systems, show that the proposed methodology allows to achieve execution times worse by only up to 8% compared to a search algorithm that performs a full search over combinations of application parameters, while taking only up to 26% time of the latter.