[R] Rmpi : openmpi and mpi.spawn.Rslaves

Paul Hewson paul.hewson at plymouth.ac.uk
Wed Nov 28 21:11:28 CET 2007


Hello,

I'm using R on a 10 blade dual quad core Rocks Cluster, and trying to use Rpmi and snow.   I basically wondered if at the moment I ought to install Rmpi against another form of mpi (not openmpi) and wondered whether anyone could pass on any experience.

I'm mainly worried about (a) the R server taking up 100% cpu time (I think this is a known issue with Rmpi and openmpi) and (b) checking that my slaves are running on separate nodes rather than all piling onto the same nodes (on the server).   Some very crude experiments in snow imply that slave processes might well be getting passed on to separate nodes beyond the server.

However, it's a little harder convincing myself which nodes are doing what.   When looking directly in Rmpi, lamhosts() reports all the hosts correctly, unfortunately mpi.universe.size reports 1 node (I think this is as documented).   More importantly, I have the issue with mpi.spawn.Rslaves() freezing.   This has been mentioned before:
http://tolstoy.newcastle.edu.au/R/e2/help/07/06/18573.html

I wondered if anyone would mind supplying pointers.   I'm specifically wondering if I need to check the version of openmpi, and whether at the moment it would be better to use another flavour of mpi.

I would be extremely grateful for any advice.

Thanks

Paul

-=-=-==-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Paul Hewson
Lecturer in Statistics
University of Plymouth
Drake Circus
Plymouth PL4 8AA

tel ++44(0)1752 232778
email paul.hewson at plymouth.ac.uk
web http://www.plymouth.ac.uk/staff/phewson



More information about the R-help mailing list