ABSTRACT
One application area that has made use of SCTP and its features has been middleware for MPI (Message Passing Interface), a library widely used in High Performance Computing to write message-passing programs for scientific computation. The group at the University of British Columbia has released modules for both major open source MPI libraries: MPICH2 from Argonne National Laboratory and Open MPI from a consortium of academic, governmental, and industrial research labs. The SCTP-based middleware modules make it possible for the wide-variety of currently
Using
available MPI programs to execute using SCTP as the underlying communication protocol, without modification to the applications themselves. The middleware can take advantage of many standard features of SCTP including multihoming as well as those in extensions such as the CMT in the FreeBSD implementation.1 Utilizing these SCTP features, this work has been able to improve the reliability and performance for clusters of commodity machines for use as compute servers. In this chapter, we describe our experiences in using SCTP and multihoming features, and discuss the design as well as some experimental results.
