File Access Prediction and Grouping

Grouping related data items has many uses in storage systems, and is a more effective use of access prediction than most proposed prefetching schemes to date. Our work attempts to provide a mechanism for constructing groupings of highly related files based on observed access patterns. Such groupings can be used for layout decisions to reduce access latencies of disk drives, or reduce the number of media changes in a tertiary storage library. Grouping information can also be used to drive the group-based management of caches, reducing demand fetches and allowing implicit prefetching.

In this talk we discuss file successor tracking, its use in file successor prediction, and its utility for the construction of highly related groups of files. We demonstrate how limiting the prediction problem to immediate successors can allow us to build very simple predictors that achieve very high performance with limited resource requirements. We then discuss how using these predictions to form file groups can result in better client and server caches.