Pre-execution Data Prefetching with Inter-thread I/O Scheduling

With the rate of computing power growing much faster than that of storage I/O access, parallel applications suffer more from I/O latency. I/O prefetching is effective in hiding I/O latency. However, existing I/O prefetching techniques are conservative and their effectiveness is limited. Recently, a more aggressive prefetching approach named pre-execution prefetching [19] has been proposed. In this paper, we first identify the drawback of this pre-execution prefetching approach, and then propose a new method to overcome the drawback by scheduling the I/O operations between the main thread and the prefetching thread. By careful I/O scheduling, our approach further extends the computation and I/O concurrency and avoids the I/O competition within one process. The results of extensive experiments, including experiments on real-life applications such as big matrix manipulation and Hill encryption, demonstrate the benefits of the proposed approach.