Improving Mpi For Linux Clusters

Greg Benson, University of San Francisco Computer Science Department

Stevenson Hall 1300
12:00 PM - 12:50 PM

The Message Passing Interface (MPI) has become a dominant programming paradigm for scientific and numerical parallel computing. In addition, Linux clusters base on switched networks allow for rapid and low-cost deployment parallel computing systems. This talk will summarize our recent efforts to improve MPI for Linux clusters. In particular, our own implementation of MPI, called USFMPI, performs better than both MPICH and LAM-MPI in many cases. Unlike MPICH and LAM-MPI, the USFMPI implementation is relatively compact making it easy to explore alternative implementation techniques. Experiments with the dissemination algorithm for allgather collective communication will be described.