Efficiency Evaluation of Cray XT Parallel IO Stack.pdfVIP

Efficiency Evaluation of Cray XT Parallel IO Stack Weikuan Yu1,2, Sarp Oral2, Jeffrey Vetter1, Richard Barrett1,2 Computer Science and Mathematics Division1 National Center for Computational Science Division2 Oak Ridge National Laboratory One Bethel Valley Road, Oak Ridge, TN 37831 {wyu,oralhs,vetter,rbarrett}@ ABSTRACT: PetaScale computing platforms need to be coupled with efficient IO subsystems that can deliver commensurate IO throughput to scientific applications. In order to gain insights into the deliverable IO efficiency on the Cray XT platform at ORNL, this paper presents an in-depth efficiency evaluation of its parallel IO software stack. Our evaluation covers the performance of a variety of parallel IO interfaces, including POSIX IO, MPI-IO, and HDF5. Moreover, we describe several tuning parameters for these interfaces and present their effectiveness in enhancing the IO efficiency. KEYWORDS: Cray XT, Parallel IO, Efficiency, ROMIO, Lustre 1 Introduction High Performance Computing (HPC) is soon approaching the era of PetaScale computing. Massive Parallel Processing (MPP) platforms such as Lawrence Livermore National Lab’s (LLNL) BlueGene/L [4] and Oak Ridge National Lab’s (ORNL) XT4 (Jaguar) [18] have already reached 100s of TFLOPS. Large-scale data-intensive scientific applications on these platforms demand increasingly higher IO throughputs as well as higher computational capacity. Correspondingly, scalable parallel IO needs to be available for these applications to perform well. In fact, aggregated IO throughput ranging from 10-100 GB/s has been reported for Red Storm and Thunderbird at Sandia National Laboratory, and Tera-10 at CEA in Europe [6], among many others. However, such performance marks are the systems’ peak IO bandwidth, which are usually measured with best suitable parameters for the IO subsystem, while what matters the most to the end users is the deliverable IO bandwidth to their respective scientific applica