Early experiences with the myricom 2000 switch on an SMP Beowulf-class cluster for unstructured adaptive meshing

We explore the current capabilities of the recently released Myricom 2000 switch, using MPICH-GM for communication, on a 2-way SMP Pentium III Beowulf-Class cluster. Performance measurements indicate that data transfer rates of approximately 225 Mbytes/s with 9.3μ seconds latency for ping-pong tests can be achieved for messages as large as 32 Mbytes. When shared-memory communication is used approximately 130 MBytes/s with 1.5μ seconds latency for long messages (250 MBytes/s peak) is possible. The performance varies depending on how processors communicate; either within an SMP node or across nodes. Performance for parallel unstructured adaptive refinement of 3D tetrahedral meshes shows noticeable improvement when compared to 100BaseT Ethernet. Furthermore, when compared to traditional systems such as the SGI Origin 2000, the combination of this fast network with high performance SMP processors demonstrate that Beowulf-Clusters compare favorably with such systems - even for communication intensive applications.