On Sat, 2002-05-11 at 22:38, Ron Chen wrote: > 3. Network > 100 Mbps may be the best, in terms of cost-effetive. > But if you run parallel BLAST, then you may consider > high-speed, low latency like Myrinet. Parallel BLAST (on Linux clusters) does not generally require such a network. There may be valid other reasons to have a low latency high bandwidth network. Parallel BLAST is generally done either using the built in pthreads support, or via coarse grain job segmentation. The pthreads mechanism depends upon the machines SMP ability and OS support for pthreads. Pthreaded programs generally do not make use of low latency interconnects unless some attempt is made to add support for MPI or other mechanism. The coarse grained mechanisms depend upon a segmentation operation, and then a data distribution. This distribution tends to be quite small (though not always). Joe