An optimized MPI library for VIA/SCI cards

Rapid developments in computer architecture and in networking technology have driven the construction of clusters of cluster. Now cluster computers are an inexpensive alternative to parallel computers. High bandwidth low latency communication networks have become more and more important. In conjunction with the Scalable Coherent Interface (SCI) the Virtual Interface Architecture (VIA) offers excellent opportunities for high speed networks. For this reason our research group is developing such a VIA/SCI network card. This paper describes the demands and the design of a new MPI library for heterogeneous communication networks. The implementation of this MPI library is demonstrated on an example VIA/SCI card.