Many applications, ranging from simple utilities to databases, manipulate data too large or infrequently used to be reliably found in memory caches. Due to the huge disparity between processor speeds and disk access times, these disk-bound applications tend to waste a disproportionate percentage of their execution times waiting for disk requests to complete. Disk latency can be hidden from these applications by taking advantage of under-utilized I/O resources to perform prefetching. To be effective, however, prefetching requires timely and accurate information about what data will be needed in the future.

In this talk, Chang will present a novel approach to automatically identifying the future data needs of applications. The key idea is to pre-execute application code during spare processing cycles, using only the available data state. Unlike previous automatic approaches, this approach can be applied to generic disk-bound applications. She will describe the design and evaluation of SpecHint, a tool that transforms application binaries to use the speculative execution approach. The evaluation demonstrates that this approach provides substantial benefits for a range of disk-bound applications in a variety of operating conditions.