| This thesis presents a comprehensive software framework—Automodeler—to provide automatic modeling and forecasting for input/output (I/O) request interarrival times. In Automodeler, ARIMA models of interarrival times are automatically identified and built during application execution. Model parameters are recursively estimated in real-time for every new request arrival, adapting to changes that are intrinsic or external to the running application. Online forecasts are subsequently generated based on the updated parameters.; Our goal is to improve application-level I/O performance by using forecasts to guide the prefetch of I/O requests before they arrive, reducing stall time. We combine a just-in-time prefetcher with ARIMA time series predictions and Markov model spatial predictions [41] to adaptively determine when, what, and how many data blocks to prefetch.; To validate our approach, we built a prototype that integrates adaptive prefetching with caching and local disk striping in the PPFS2 [51] testbed. Results obtained for a computational physics code demonstrate 30% improvement in total execution time over the traditional Unix file system on three Linux clusters, equipped with different hardware configurations. More importantly, this performance improvement has small memory requirements and is shown to scale with increasing I/O intensity. |
