Probabilistic-Based Selection of Alternate Implementations for Heterogeneous Platforms

Abstract

Over the last years, heterogeneous architectures have become a de facto approach for improving the performance of numerous scientific and industrial applications. However, developing for these architectures is not straightforward: each processor demands its specific programming paradigm and, often, certain applications are only well-suited to run on a particular processing unit. Therefore, a major challenge arises when programming for these platforms: to select the most suitable device and routine implementation to solve a given problem. To deal with this issue, this paper proposes a novel probabilistic-based selector that uses the problem size to automatically choose the most appropriate version of a same kernel. In order to analyze this approach, we have developed this selector within the OmpSs programming framework and evaluated its accuracy and performance gains when executing different implementations of the general matrix-matrix multiplication. Finally, we also demonstrate how this solution delivers a comparable performance with respect to a runtime approach from the state-of-the-art.

J. Fernández—This work was partially supported by the EU project ICT 644235 “RePhrase: REfactoring Parallel Heterogeneous Resource-Aware Applications” and the project TIN2013-41350-P “Scalable Data Management Techniques for High-End Computing Systems” from the Ministerio de Economía y Competitividad, Spain.