Comparing MPI Performance of SCI and VIA

Both the Scalable Coherent Interface (SCI) and the Virtual Interface Architecture (VIA) aim at providing effective cluster communication. While the former is a standardized subnet technology the latter is a generic architecture which can be applied to a variety of physical medias. Both approaches provide user level communication, but they achieve it on different ways and thus, have different characteristics that are independent of the actual implementation. In this paper we report and compare the raw network speed of an SCI and a VIA implementation as delivered by MPI and show how it affects application performance by means of the NAS Parallel Benchmark Suite. Keywords—NetPIPE, NAS Parallel Benchmarks, SCI, VIA, MPI