The Design, Implementation, and Evaluation of a Predictive Caching File System

This paper describes the design, implementation, and evaluation of a predictive file caching approach. Predictive caching reduces file system latency by converting the file cache into a staging area for data about to be referenced as opposed to the conventional repository of recently accessed data. The system uses heuristic-based algorithms to predict and prefetch data without user or application intervention. A pruning algorithm keeps prefetching overhead in check. We show the expected performance gains for a variety of simulated environments. Simulations run on file traces show that prefetching can improve cache miss rates by as much as 70% and read times by as much as 52%. For certain common workloads (e.g., compilation), our algorithm achieves performance similar to that of an optimal prediction system. Our prototype SunOS implementation verifies the simulation results and shows read improvements of as much as 61% for real applications. Finally, our results show that prefetching can help small memory machines achieve the same performance as machines with twice as much memory