I/O characterization on a parallel file system

In this paper we present a study of I/O access patterns of scientific and general applications on a parallel file system. Understanding I/O access patterns is an essential condition to effectively designing a file system. Supercomputing applications running on these parallel systems make extensive use of parallel file systems taking advantage of faster data access by requesting information from multiple nodes simultaneously. However, parallel file systems can become a bottleneck if the file distribution parameters do not fit the access scheme of the applications. In our work, we examine a variety of such applications, providing measurement of inter-arrival times, I/O request size and burstiness demanded from a parallel file system. Our tests were conducted on the open source PVFS parallel file system with different configurations of metadata servers and I/O nodes. Among the findings are that the standard assumption of Poisson or random interarrival times is not justified and that access sizes are smaller than would be expected for a parallel application.