high performance computing on graphics processing units: hgpu.org

GPU-accelerated data-intensive applications demonstrate in excess of ten-fold speedups over CPU-only approaches. However, file-driven data movement between the CPU and the GPU can degrade performance and energy efficiencies by an order of magnitude as a result of traditional storage latency and ineffectual memory management. In this paper, we first analyze these two critical performance bottlenecks in GPU-accelerated data processing. We then study design considerations to reduce the overheads imposed by file-driven data movements in GPU computing. To address these issues, we prototype a low cost and low power all-flash array designed specifically for stream-based, I/O-rich workloads inherent in GPUs. As preliminary evaluation results, we demonstrate that our early-stage all-flash array solution can eliminate 60% ~ 90% performance discrepancy between memory-level GPU data transfer rates and storage access bandwidth by removing unnecessary data copies, memory management, and user/kernel-mode switching in the current system software stack. In addition, our all-flash array prototype consumes less dynamic power than the baseline storage array by 49%, on average.