Dynamic Kernel/Device Mapping Strategies for GPU-Assisted HPC Systems

With their high computation throughput and outstanding performance-per-watt figures, the graphics processing units (GPU) are becoming increasingly important for high-performance computing (HPC) systems. Existing GPU execution environment restricts the GPU usage to local host node. This is suitable for standalone computer nodes, but becomes inefficient for HPC systems that consist of a large number of GPU-assisted nodes. In this paper, a novel framework is proposed to support dynamic GPU kernel/device mapping strategies for HPC systems. Adaptive mapping policies are designed to mitigate the impact of network transfer overhead. The performance of the framework is studied through extensive simulations. The results show that compared with existing local-only static mapping method, the proposed framework is capable of improving the system-wide GPU utilization rate and computation throughput, especially when the concurrent workloads exhibit different GPU usage intensities.