SFDC: File Access Pattern Aware Cache Framework for High-performance Computer

摘要

The performance gap between computation and storage in HPC systems is enlarged with the rapid increase of compute scale. Client-side file caching has been proved to be an effective technique to improve I/O performance in parallel file system adopted by HPC systems. Current cache frameworks mainly focus on local thread access frequency, which suffers from cache inefficiency problems for the lack of file level access information. This paper presents SFDC, a caching framework that aims to improve I/O performance through providing high cache efficiency by using file level access knowledge. SFDC collects access traces of each file from all the processes, evaluates each file according to access traces, evicts valueless file data from cache space as early as possible and prefetches valuable file data before being used. Moreover, SFDC can reduce small access to underlying parallel file system and conduct file stage before application running, further improve I/O performance. Performance evaluation is conducted with IOR benchmark and a real application called SUM on TH-1A supercomputer. The results show significant performance enhancement.