Topology-Aware Data Aggregation for High Performance Collective MPI-IO on a Multi-core Cluster System

Abstract

Parallel I/O such as MPI-IO is one of the performance improvement solutions in parallel computing using MPI. ROMIO is a widely used MPI-IO implementation which addresses to improve collective I/O performance by using its optimization named two-phase I/O. File I/O task is given to a subset of or all of MPI processes, which are called aggregators. Multiple CPUs or CPU cores give a chance to increase computing power by deploying multiple MPI processes per compute node, while such deployment leads to poor I/O performance due to ROMIO's topology-unaware aggregator layout. In our previous work, optimized aggregator layout which was suitable for striping accesses on a Lustre file system improved I/O performance, however, its unbalanced communication load due to unawareness in MPI rank layout among compute nodes led to ineffective data aggregation. To address minimization in data aggregation time for further I/O performance improvements, we introduce a topology-aware data aggregation scheme which takes care of MPI rank layout across compute nodes. The proposal arranges data collection sequence by aggregators in order to mitigate network contention. The optimization has achieved up to 67% improvements in I/O performance compared with the original ROMIO in HPIO benchmark runs using 768 processes on 64 compute nodes of the TSUBAME2.5 supercomputer at the Tokyo Institute of Technology. Even if the number of aggregators was half or 1/3 of the total number of processes, the optimization has still kept comparable I/O performance with the maximum performance.