Performance comparison of LAM/MPI, MPICH, and MVICH on a linux cluster connected by a gigabit ethernet network

Abstract

We evaluate and compare the performance of LAM,
\nMPICH, and MVICH on a Linux cluster connected
\nby a Gigabit Ethernet network. Performance statistics
\nare collected using NetPIPE which show the
\nbehavior of LAM/MPI and MPICH over a gigabit
\nnetwork. Since LAM and MPICH use the TCP/IP
\nsocket interface for communicating messages, it is
\ncritical to have high TCP/IP performance. Despite
\nmany eorts to improve TCP/IP performance,
\nthe performance graphs presented here indicate that
\nthe overhead incurred in protocol stack processing
\nis still high. Recent developments such as the
\nVIA-based MPI implementation MVICH can improve
\ncommunication throughput and thus give the
\npromise of enabling distributed applications to improve
\nperformance.