In this study, we designed a caching parallel file system for node-local persistent memory. The caching parallel file system improves access performance to the backend parallel file system in a way transparent to users. Although existing caching parallel file systems succeed in improving bandwidth, the performance of metadata is limited by that of the backend parallel file system. Based on the CHFS ad hoc parallel and distributed file system, our system maintains the high metadata performance of CHFS by relaxing the consistency with the backend system in easy-to-understand semantics. Upon implementation, our system's performance was evaluated using the IO500 benchmark, and was found to exhibit almost identical performance to that of the original CHFS. This shows that it is possible to realize a caching parallel file system without sacrificing metadata performance.
Bio : Osamu Tatebe received a Ph.D. degree in computer science (1997, Univ. of Tokyo). He worked at Electrotechnical Laboratory (ETL), and National Institute of Advanced Industrial Science and Technology (AIST) until 2006. He is now a professor of Center for Computational Sciences at University of Tsukuba. His research area is high-performance computing, data-intensive computing, and parallel and distributed system software.