ETCG: Energy-Aware CPU Thread Throttling for CPU-GPU Collaborative Environments

Abstract

High-Performance computing systems have been constantly adopting CPU-GPU architectures as a collaborative environment to accelerate applications by partitioning threads/kernels execution across both devices. However, exploiting the synergetic benefits of this system is challenging, since maximizing resource utilization by triggering the highest number threads is not always the best strategy to optimize performance or energy consumption. This work shows that selecting the right number of CPU threads in a CPU-GPU collaborative environment is even trickier. To address this problem, we propose ETCG - Energy-aware CPU Thread throttling for CPU-GPU collaborative environments. ETCG transparently selects a near-optimal number of CPU threads to minimize the energy-delay product (EDP) of CPU-GPU applications. Compared to the use of the maximum number of threads supported by the hardware, ETCG provides, on average, 73% of EDP reduction. In addition, ETCG shows, on average, 3% less EDP by just taking 5% of searching time compared to the optimal solution.