Kate Ericson, Colorado State University

Searching multidimensional parameter spaces to optimize Parallel I/O

Abstract: We describe efforts to optimize the Parallel I/O (PIO) library, an MPI-IO based I/O library used in version 4 of the Community Climate System Model (CCSM). We use two methods to explore the large parameter space with PIO: Active Harmony, which uses a simplex method, and simulated annealing. We report on results optimizing disk I/O bandwidths on both a Blue Gene/L at NCAR and a Cray XT5 at NICS. Preliminary results on the Blue Gene/L indicate that we sustain 70-90% of the peak disk I/O bandwidth.

Presentation slides (pdf)