Performance optimization for CPU-GPU heterogeneous parallel system

Abstract

With GPU (Graphics Processing Unit) taking part in general-purpose computing, a heterogeneous system usually achieves higher performance and efficiency. There are many studies on how to improve the performance of a heterogeneous system, among of which are a number of researches to achieve the goal by allocating workload into processors with different strategies. In the paper, we implement a task allocation model in the principle of making execution time of the partition on CPU closer to the partition on GPU to the maximum extent. The task allocation process contains two stages. Firstly, we make use of SVM (Support Vector Machine) to classify the tasks into two sets as CPU-kind and GPU-kind in pre-treating stage. Secondly, we adjust the two task sets in the light of the characteristic and current running status of processors, then we map the two well-adjusted task sets to processors. Moreover, we evaluate the proposed model by implementing them on a real heterogeneous system and several benchmarks. Experimental results demonstrate that our model can achieve up to 23.43% of performance improvement compared to some states of the art allocation strategies averagely.