Grid Enabled MPI:PACX-MPI Optimization

Abstract

As the power of computer progresses, software get increasingly more complex and intelligent. Meanwhile, network infrastructure has also imporved at a high speed. Because of these advancements, there is a strong interest in sharing computing resources scattered over many different places. Therefore, the large-scale resources sharing and management become very important issues. At the beginning, the sharing of computational resources is the primary interest of resource sharing. A typical case is the so-called cluster of cluster(CoC), which consists of a group of clusters working together to solve a large computation-intensive application. To facilitate MPI programs to function properly across clusters, MPI extensions for parallel system, such as MPICH-Globus2 and MPICH-VMI have been introduced and are briefly discussed here. We also discuss how these MPI implementations work across clusters behind firewalls. In addition, an optimization of PACX-MPI for data transmission between two clusters has been implemented. Communications between two clusters are studied and handled with multithreading. System throughputs are measured by use of NAS Parallel Benchmark[16] and Persistence of Vision Raytracer[17].