e-CLAS: Effective GPUDirect I/O Classification Scheme

GPUs are widely used in various computational areas, due to their massive multithreading platform capable of high-level parallelism. Especially, to speedup I/O processing, GPUDirect I/O has been developed to enable direct data access between GPU memory and either NVMe or NVMe-oF. However, GPUDirect I/O does not always enhance the I/O performance of applications, because of reducing the benefit of page cache. In this paper, we present an effective I/O scheme, called e-CLAS, that can categorize I/Os through machine learning algorithms, to maximize the benefit of GPUDirect I/O while managing NVMe device storage at cost-effective way. The performance evaluation shows that our scheme effectively works to increase I/O performance, by appropriately classifying I/Os generated from applications.