SBIR Phase II: MPI-2: A Systematic Study, Design, and Commercialization of the Extended Message Passing Interface for NOWs and Parallel Computers

Mpi Software Technology, Inc., Birmingham AL

Investigators

Abstract

This Small Business Innovation Research Phase II project builds on MPI Software Technology, Inc.'s (MSTI's) extensive MPI-1 implementation experience with our MPI-2 design developed in Phase I in order to continue research and advanced prototyping of a quality implementation of the MPI-2 standard for clusters of workstations . In Phase II, we build upon prototypes created in Phase I, while continuing our investigations into scalable dynamic process startup, advanced and poly-algorithmic approaches to one-sided communications, collective operations, and parallel I/O. Our research and development outcomes will enable the high-performance computing community to unlock the potential of the latest workstation and networking technology, providing access to architectural enhancements of systems and software, and more complex computational environments. Rationale for undertaking this effort is that the scientific community needs enhancements to its most important parallel processing environment, MPI, and that workstation cluster targets comprise the fastest growing component of parallel processing environments. Inventing MPI-2 capability for HPC represents widely enabling technology for scientists and engineers to produce new science, while incorporating computer-science challenges of its own, both of a research and advanced development nature. Significant software, protocol, and algorithmic challenges must be tackled in order to create a useful MPI-2 environment. MPI-2's dynamic process management would support several classes of new scientific applications, and computation strategies. These include computational servers, growing/shrinking parallel applications, and multi-disciplinary codes. Support for the one-sided model would enable classes of applications that need fine grain communication support, including certain sparse matrix algorithms, as well as quantum chemistry codes that utilize global array-type algorithms (e.g., Focker-Planck computations) . Support for effective intercommunicator collective operations would simplify and enable applications that work with dataflow models, including composite parallel simulation and visualization techniques. Support for MPI-2's I/O techniques would support myriad out-of-core and database-type applications, including growing interest in financial modeling with MPI, but also the traditional scientific problems in areas of climate and weather modeling, and others with large, out-of-core datasets needed in conjunction with parallel computing.