MPI-2 One-Sided Communications on a Giganet SMP Cluster
We describe and evaluate an implementation of the MPI-2 one-sided communications on a Giganet SMP Cluster. The cluster runs under Linux with our own port of MPICH 1.2.1, a well-known, portable, non-threaded MPI implementation, to the Virtual Interface Provider Library (VIPL). We call this implementation MPI/VIA. The one-sided communications part was adapted from MPI/SX, a full MPI-2 implementation for the NEC SX-5 vector-parallel supercomputer. We evaluate the performance of the one-sided communications by comparing to point-to-point communication. For applications that can be implemented naturally in both the one-sided and the point-to-point model, the MPI/VIA implementation is such that the user is not penalized by choosing one model over the other.
KeywordsControl Message Target Window Synchronization Overhead Communication Request Remote Memory Access
Unable to display preview. Download preview PDF.
- 1.N. Asai, T. Kentemich, and P. Lagier. MPI-2 implementation on Fujitsu generic message passing kernel. In Supercomputing, 1999. http://www.sc99.org/proceedings/papers/lagier.pdf.
- 2.S. Booth and E. Mourão. Single sided MPI implementations for SUN MPI. In Supercomputing, 2000. http://www.sc2000.org/proceedings/techpapr/index.htm/#01.
- 3.Compaq, Intel, Microsoft. Virtual Interface Specification, 1997.Google Scholar
- 4.W. Gropp, S. Huss-Lederman, A. Lumsdaine, E. Lusk, B. Nitzberg, W. Saphir, and M. Snir. MPI-The Complete Reference, volume 2, The MPI Extensions. MIT Press, 1998.Google Scholar
- 6.J. L. Träff, H. Ritzdorf, and R. Hempel. The implementation of MPI-2 one-sided communication for the NEC SX-5. In Supercomputing, 2000. http://www.sc2000.org/proceedings/techpapr/index.htm/#01.