Nick Jones, University of Tennessee, Knoxville

Characterizing the Performance of Kraken's Lustre File System

Abstract: Application I/O performance of the lustre filesystem on the National Institute for Computational Science (NICS) Cray XT5 can vary significantly. For example we have observed that the I/O performance of the high-resolution Community Climate System Model (CCSM) can vary from 5x to 18x. We have developed a set of tools for analyzing and visualizing the I/O performance of large parallel computations at NICS. In particular we analyze the impact job placement has on network congestion and read and write bandwidth to the lustre filesystem.

Presentation slides (mov)