Energy Efficient Dynamic Load Balancing over MultiGPU Heterogeneous Systems

Current HPC technologies demand high amounts of power and energy to achieve good performances. In order to address the next milestone in peak power, powerful graphic processing units and manycore processors present in current HPC systems need to be programmed having energy efficiency in mind. As energy efficiency is a major issue in this area, the existing codes and libraries need to be adapted to improve the use of the available resources. Rewriting the code requires deep knowledge of programming and architectural details to achieve good efficiency, which is an ad-hoc solution for a concrete system. We present the Ull_Multiobjective_Framework, an interface that allows automatic dynamic balance of the workload for parallel iterative codes in heterogeneous environments. UllMF allows to include the overall energy consumption as a parameter during the balancing process. This tool hides all architectural measurement details, requires very low effort to the programmer and introduces a minimum overhead. The calibration library has been used to solve iterative problems over heterogeneous platforms. To validate it, we present an analysis of different Dynamic Programming problems over different hardware configurations of a MultiGPU heterogeneous system.