stephen mulcahy wrote:
I'm still unsure what the bottleneck currently is. My hunch is that a faster interconnect *should* give a better performance but I'm not sure how to quantify that. Do others here running MPI jobs see big improvements in using Infiniband over Gigabit for MPI jobs or does it really depend on the characteristics of the MPI job? What characteristics should I be looking for?
Long time ago I was working with some guys who were using Dimemas (http://www.cepba.upc.es/dimemas/). This tool allows you to simulate the app on any network.
You first need to run the app once to collect a trace of all MPI-calls that are performed during a run. Next, the tool allows you to simulate your app (based on the trace to the MPI-calls) on any network (taking into account how long the communication takes on the network you wanted to simulate).
I can imagine this tool might be able to simulate your app on other types of networks and investigate the speedup you might obtain.
toon disclaimer: I have no affiliation with the University of Barcelona. _______________________________________________ Beowulf mailing list, [email protected] To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
