Assessing Energy Consumption and Runtime Efficiency of Master- Worker Parallel Evolutionary Algorithms in CPU-GPU Systems

Abstract

Thanks to parallel processing, it is possible not only to reduce code runtime but also energy consumption once the workload has been adequately distributed among the available cores. The current availability of heterogeneous architectures including GPU and CPU cores with different power-performance characteristics and mechanisms for dynamic voltage and frequency scaling does, in fact, pose a new challenge for developing efficient parallel codes that take into account both the achieved speedup and the energy consumed. This paper analyses the energy consumption and runtime behavior of a parallel master-worker evolutionary algorithm according to the workload distribution between GPU and CPU cores and their operation frequencies. It also proposes a model that has been fitted using multiple linear regression and which enables a workload distribution that considers both runtime and energy consumption by means of a cost function that suitably weights both objectives. Since many useful bioinformatics and data mining applications are tackled by programs with a similar profile to that of the parallel master-worker procedure considered here, the proposed energy-aware approach could be applied in many different situations.