HPSM: A Programming Framework for Multi-CPU and Multi-GPU Systems

This paper presents a high-level C++ framework to explore multi-CPU and multi-GPU systems called HPSM. HPSM enables parallel loops and reductions implemented over three parallel backends: Serial, OpenMP (with GCC and libKOMP runtime), and StarPU. We evaluated HPSM development effort with AXPY program, and performance with three parallel benchmarks: N-Body, Hotspot, and CFD solver. The CPU-GPU combination attained better performance than only GPUs for all cases on a CPU-GPU system. Still, our findings provide evidence that NUMA affinity at framework level may produce different results.