Optimization of Data-Parallel Applications on Heterogeneous HPC Platforms for Dynamic Energy Through Workload Distribution

Abstract

Energy is one of the most important objectives for optimization on modern heterogeneous high performance computing (HPC) platforms. The tight integration of multicore CPUs with accelerators in these platforms present several challenges to optimization of multithreaded data-parallel applications for dynamic energy. In this work, we formulate the optimization problem of data-parallel applications on heterogeneous HPC platforms for dynamic energy through workload distribution. We propose a solution method to solve the problem. It consists of a data-partitioning algorithm that employs load imbalancing technique to determine the workload distribution minimizing the dynamic energy consumption of the parallel execution of an application. The inputs to the algorithm are discrete dynamic energy profiles of individual computing devices. We experimentally analyse the proposed algorithm using two multithreaded data-parallel applications, matrix multiplication and 2D fast Fourier transform. The load-imbalanced solutions provided by the algorithm achieve significant dynamic energy reductions (on the average 130% and 44%) compared to the load-balanced ones for the applications.