Performance Modeling and Evaluation of MPI-I/O on a Cluster

摘要

Cluster computing is an area of growing interest in search to support parallel and distributed applications. Many of these applications are I/O intensive, and the limited bandwidth of the I/O subsystem of the cluster is an important bottleneck that is usually ignored. Thus, the performance of parallel I/O primitives is critical for overall cluster performance. In this work, we characterize the performance of basic ROMIO MPI-I/O routines on a PC cluster using the NFS and PVFS file systems. Our goal is to detect weak spots in the use of these routines and to predict their impact on the application's performance.