A Decoupled Architecture for Application-Specific File Prefetching

Abstract

Data-intensive applications such as multimedia and data mining programs may exhibit sophisticated access patterns that are difficult to predict from past reference history and are different from one application to another. This paper presents the design, implementation, and evaluation of an automatic application-specific file prefetching (AASFP) mechanism that is designed to improve the disk I/O performance of application programs with such complicated access patterns. The key idea of AASFP is to convert an application into two threads: a computation thread, which is the original program containing both computation and disk I/O, and a prefetch thread, which contains all the instructions in the original program that are related to disk accesses. At run time, the prefetch thread is scheduled to run suciently far ahead of the computation thread, so that disk blocks can be prefetched and put in the file buffer cache before the computation thread needs them. Through a source-to-source translator, the conversion of a given application into two such threads is made completely automatic. Measurements on an initial AASFP prototype under Linux show that it provides as much as 54% overall performance improvement for a volume visualization application.