MPI Implementation

SCS has developed an MPI/FCI* library for its TNet hardware. Using this library for parallel applications achieves communication latencies below 13.5 us and a process-to-process bandwidth of 100 MB/s. This performance is achieved with an operating system bypass - we use a direct communication channel between each application memory.

The MPI/FCI library fulfills MPI standard 1.2. It passes the INTEL/DARPA test-suite - over 300 different tests covering the whole MPI functionality. The MPI/FCI is available for COMPAQ Tru64 Unix, as well as for Linux on Intel and Alpha platforms.

* Message Parsing Interface - MPI (http://www.mpi-forum.org/) is a standardized communication library for massive parallel computer systems. It is used in loosly-coupled workstation clusters (e.g. Beowulf systems) - up to the world’s biggest special-purpose machines such as the ASCI Red.


Download here the complet project-description as pdf-file.