Abstract
Input/output (I/O) operations can represent a significant proportion of the run-time when large scientific applications are run in parallel. Although there have been advances in the form of file-format libraries, file-system design and I/O hardware, a growing divergence exists between the performance of parallel file-systems and compute processing rates. The effect is often a bottleneck when any form of file-system interaction is required. In this paper we present RIOT – an input/output tracing toolkit being developed at the University of Warwick for dynamic attachment to parallel applications. The two-stage tracing process includes a lightweight library to record I/O events and an in-depth post-execution analysis tool to extract performance metrics such as MPI-IO bandwidth, effective POSIX/file-system bandwidth, duration of individual or aggregated time spent in obtaining or releasing file locks and temporal information relating to parallel file activity. We present a case study on the use of RIOT for three standard industry I/O benchmarks: the BT-IO micro-application from NASA’s Parallel Benchmark suite, FLASH-IO, a benchmark which replicates the check-pointing operations of the FLASH thermonuclear star modelling code and IOR, an industry standard I/O benchmark using HDF-5 and MPI-IO. Furthermore, we utilise RIOT to assess these codes when running with the Parallel Log-structured File System (PLFS) middleware developed by the Los Alamos National Laboratory.