Evaluating transport level protocols for MPI in the Internet

摘要

There is increased interest in executing parallel programs in wide area networks and the Internet. The de-facto standard transport protocol for the Internet is TCP and the de-facto communication library for distributed parallel programs is MPI, In this paper we evaluate the performance of using TCP and a public domain version of MPI (LAM-MPI) in an emulated wide area network environment. We show how performance is affected by latency and loss. We evaluate performance with respect to TCP New-Reno and TCP SACK and also evaluate this with respect to Nagle's algorithm and delayed acknowledgments.