Toward Optimizing File IO on GPU Clusters

In general, GPU memory is independent of CPU memory, and when an application running on a GPU performs file operations, data must be temporarily copied to and from CPU memory. On the other hand, NVIDIA has recently introduced GPUDirect Storage (GDS), which enables file IO directly from the GPU. In this study, we benchmarked the file IO performance of different transfer methods, including GDS, by varying block size and other parameters. From the evaluation results, we found that GDS performs up to 1.4 times better than non-GDS for sequential access in certain measurement environments. However, non-GDS sometimes shows higher performance in some environments, so GDS did not necessarily have an advantage depending on the measurement environment or access pattern. We plan to set several parameters based on data obtained from the performance evaluations and realize these features through implementation in the HDF5 virtual file driver "vfd-gds".