> I've seen architectures with two network switchs, one is used for I/O > (writing, reading, so on) and another for message passing (MPI). how is > this achieved? I get the idea, from one place, where the applications > running must be aware of this but I was thinking that for this to work it > must be transparent to the application. How can this be achieved?
in short: you assign a different IP to each interface (as normal). it's convenient to use one non-low-order in the IP to distinguish, and also nice to have separate hostnames like node1 and node1-mpi. I wonder whether anyone has critically evaluated whether this is important. cluster people I talk to like to say fuzzy things like "separate networks make the cluster breathe better". as much as I admire car analogies, I observe that when apps are doing IO, they tend not to be doing MPI. if your workload is like that, bonding rather than partitioning would actually improve performance. I wonder whether the partitioning approach might actual reflect other constraints, such as using half-duplex hubs, or low-bisection networks. regards, mark hahn. _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
