Feedback Control Optimization for Performance and Energy Efficiency on CPU-GPU Heterogeneous Systems

Owing to the rising awareness of environment protection, high performance is not the only aim in system design, energy efficiency has increasingly become an important goal. In accordance with this goal, heterogeneous systems which are more efficient than CPU-based homogeneous systems, and occupying a growing proportion in the Top500 and the Green500 lists. Nevertheless, heterogeneous system design being more complex presents greater challenges in achieving a good tradeoff between performance and energy efficiency for applications running on such systems. To address the performance energy tradeoff issue in CPU-GPU heterogeneous systems, we propose a novel feedback control optimization (FCO) method that alternates between frequency scaling of device and division of kernel workload between CPU and GPU. Given a kernel and a workload division, frequency scaling involves finding near-optimal core frequency of the CPU and of the GPU. Further, an iterative algorithm is proposed for finding a near-optimal workload division that balance workload between CPU and GPU at a frequency that was optimal for the previous workload division. The frequency scaling phase and workload division phase are alternatively performed until the proposed FCO method converges and finds a configuration including core frequency for CPU, core frequency for GPU, and the workload division. Experiments show that compared with the state-of-the-art GreenGPU method, performance can be improved by 7.9%, while energy consumption can be reduced by 4.16%.