ClusterFetch: A lightweight prefetcher that responds to intensive disk read patterns

Abstract

Application launch and loading times are important determinants of user experience in the personal computing environment. Since these delays largely depend on the performance of secondary storage, they can be reduced by prefetching disk blocks. However, existing prefetching schemes for general workloads incur a significant overhead in analyzing correlations between blocks so as to choose the blocks to prefetch, and, more significantly, these analyses lack accuracy. We propose a lightweight prefetcher called ClusterFetch which records the sequences of I/O requests that are triggered by file requests during launch and loading. When the same application is run again, the disk blocks that correspond to the stored sequences of I/O requests are prefetched when the related files are opened. Experimental results show that ClusterFetch can reduce application launch times by up to 30.9%, and loading times by up to 15.9%.