Optimal balance between energy and performance in hybrid computing applications

The latest top 10 supercomputers are dominated by heterogeneous systems with CPUs and accelerators (GPU or Xeon Phi) tightly coupled together. As the heterogeneity of future high performance computing systems keeps increasing, it becomes paramount to judiciously use CPUs and accelerators to improve performance and/or reduce energy consumption. The widely used programming model today is to offload computation intensive workload to accelerators. Theoretically, the hybrid computing model (i.e. running a subset of calculations concurrently on both CPUs and accelerators) can potentially offer advantages of improved energy efficiency and performance. However, this is not yet a common practice due to the uncertainty of energy/performance benefits as well as the increased programming complexity. In this paper, we conduct a comprehensive study on achieving the balance between energy and performance of hybrid computing applications. We show that performance and energy optimization can be conflicting goals, the sweet spot between performance and energy consumption varies with application characteristics and is highly dependent on specific implementations, that the choice of compiler can not only influence runtime but also energy use, and that the choice of cross platform strategies (e.g. OpenCL) can result in degraded performance and increased energy.