Rmpi working with OpenMPI and PBSPro but snow fails

On Wed, 4 Mar 2009, Huw Lynes wrote:

Hi Luke,

Thanks for the quick response.

Moving onto snow in the same environment trying to setup by using
getMPICluster() returns an error in checkCluster() saying that there is
something wrong with the cluster.

I don't know what "Moving to snow" means exactly as you don't give
details of you you are starting things up so I have to guess.  If you
are using mpiexec then you need to run snow via the RMPISNOW shell
script, which for NPROCS sets up a master and a cluster with NPROCS -
1 workers, and then use

cl <- makeCluster()

to access the already running cluster.

If I take the following trivial R script:

------------------------------------------------------------------------

library(Rmpi)
library(snow)

cl <- makeCluster()
clusterCall(cl, function() Sys.info()[c("nodename","machine")])
stopCluster(cl)
------------------------------------------------------------------------

and run it as 
------------------------------------------------------------------------
#!/bin/bash
#PBS -q SMP_queue
#PBS -l select=1:ncpus=4:mpiprocs=4
#PBS -l place=scatter:excl


module load apps/R
module load libs/R-mpi

cd $PBS_O_WORKDIR
cat $PBS_NODEFILE

mpiexec RMPISNOW -f snowtest_solo.r
-----------------------------------------------------------------------

all the R processes just sit there spinning rather than doing anything
useful and I have to kill the job.

the suggestion in this mail:
https://stat.ethz.ch/pipermail/r-sig-hpc/2009-January/000069.html

results in the same problem of R spinning. I suspect that there is
something different about my OpenMPI setup that means snow is failing to
set up a master process. So you end up with all four processes as slaves
spinning on a network poll.

Huw Lynes                       | Advanced Research Computing
HEC Sysadmin                    | Cardiff University
                                | Redwood Building, 
Tel: +44 (0) 29208 70626        | King Edward VII Avenue, CF10 3NB