Selected Benchmark results obtained on the IBM cluster


The following graph illustrates the performance of the ethernet interfaces on the cluster. You can expect peak ethernet performance to be about 90Mbps for the fast-ethernet and ~800Mbps for the Gigabit Ethernet. Latency is similar for both at around 110-130 milliseconds. Note that you must use Jumbo Frames (mtu larger than 1500 bytes) to achieve throughput above 400Mbps. However, at 800Mbps (=100MB/sec) the performance of the network is substantially greater than disk IO.

Note that MPI does not fully realize the gain in performance offered by Jumbo Frames. Also there is a buffer problem around 4KB in the each of the MPI curves. Finally, note that the Alteon switch delivers almost the full back-back performance and is thus not an issue in the overall performance.