Tuning an application's I/O and tuning for compute or memory share similar characteristics, but I/O tuning introduces some unique challenges. We will examine a typical high performance computing (HPC) system and its storage. Through benchmarks, I/O kernels, and characterization tools we will build up the kind of mental model of a system's I/O performance that can help us find solutions to I/O performance problems when we encounter them.
We've given a high-level survey of parallel I/O at SC for many years. We've also had a chance to go into a little more detail at the ATPESC training program's "Data and I/O" day. In this half-day workshop, however, we are going to have a chance to go into more detail about tuning parameters and workloads. We will also cover some of the tools and resources we can use to understand why we see the performance we do.
Speaker Bio: Rob Latham is a 2022 BSSW Fellow. He strives to make scientific applications use I/O more efficiently. After earning his BS (1999) and MS (2000) in Computer Engineering at Lehigh University (Bethlehem, PA), he eventually ended up at Argonne National Laboratory (ANL). His research focus has been on high performance I/O for scientific applications and I/O metrics. He has worked on the ROMIO MPI-IO implementation, the parallel file systems PVFS (v1 and v2), Parallel NetCDF, and Mochi I/O services.
Join ZoomGov Meeting