[R-sig-Debian] rmpi/snow grabs all available CPU
Faheem Mitha
faheem at email.unc.edu
Thu May 8 20:04:57 CEST 2008
On Thu, 8 May 2008, Dirk Eddelbuettel wrote:
>
> Faheem,
>
> There is another difference.
>
> Under LAM, you could start R rightaway and it would find the hosts for MPI on
> the net.
>
> Under OpenMPI, I do something like
>
> orterun --hostfile ~/.my-ompi-hosts -np 1 R --no-save
>
> and then use
>
> cl <- makeCluster(spec=N, type="MPI")
>
> to start 'N' nodes. Note how orterun is instrumented to use just one
> 'outer' R node (our master) which will then spawn the 'inner' N nodes.
>
> If I launch things this way things behave better. I'd invite other to share
> their tips and tricks :)
>
> Hth, Dirk
Hi Dirk,
Thanks for the tips. I'll give it a try, but it seems likely that moving
back to lam might be the simplest route, till the openmpi people get their
stuff sorted out. I hope they reconsider the whole "CPUs running at full
tilt" thing. Regardless what they say, I think it is obnoxious and
unnecessary (since none of the other MPI implementations do this)
behavior.
Take care, Faheem.
More information about the R-SIG-Debian
mailing list