Asynchronous I/O Optimization for X-Ray Imaging via GPUDirect Storage

Du Wu, Peng Chen, Yiyu Tan, Yusuke Tanimura, Toshio Endo, Satoshi Matsuoka, Mohamed Wahib

Abstract: GPUDirect Storage, a novel tool provided by Nvidia, facilitates better utilization of GPU I/O by avoiding extra copies through a bounce buffer in the CPU host memory and enabling direct memory access. This technology offers significant advantages, particularly its high throughput capabilities and low latency. However, it also presents challenges in implementation due to strict layout requirements. In the case of distributed image reconstruction, data transfer often lack page alignment, which prevents utilization of GPUDirect Storage. This paper proposes a novel CPU-GPU coordination strategy to eliminate the strict page-alignment constraints, enabling arbitrary GPUDirect Storage Asynchronous I/O from or to SSDs.