[R-sig-hpc] MPICH2 best practices?
ostrouchovg at ornl.gov
Thu Apr 18 17:26:19 CEST 2013
Both OpenMPI and MPICH are implementations of the MPI standard for
cluster communication. There are also vendor implementations, such as
SGI's MPT. Both Rmpi and pbdMPI (and rest of pbdR packages) work with
either implementation. From inside R, there is no syntactical difference
between the implementations.
MPI is really intended for batch runs on a cluster. That said, a few R
packages (including snow and Rmpi) have tools to run MPI interactively
in a manager-workers style by keeping scripts active on the workers.
If you are willing to run in batch with "mpirun Rscript ...", consider
the pbdDEMO package as a way to get started through examples. In my
opinion this is a much simpler way to do parallel computing on clusters.
It is also the way MPI is used by the mainstream parallel computing
community. You can also use the parallel package (it's multicore side)
within MPI for a second level of parallelism.
On 4/18/13 10:34 AM, Rodney Sparapani wrote:
> Hi Gang:
> We recently have created an MPI cluster of 4 Linux servers running
> MPICH2 v1.2.1. I have tested MPI and it is working. I have been using
> the parallel package in my scripts up to now. But, I would like to
> switch to something that will allow me to run my R scripts on the new
> MPI cluster.
> I have managed to install Rmpi and I have been studying it.
> However, I am wondering if Rmpi is the right way to go. Rmpi seems
> to assume OpenMPI, but I don't know if that matters. I've never
> used OpenMPI so I don't know if it is really that different: it
> is still MPI so it can't be that different, right? I've been looking
> at other packages too like snow et al.
> It's a lot to digest and I am having dyspepsia ;o) Anyways, any
> advice, tips, etc. for an MPI newbie who uses R are very welcome.
George Ostrouchov, Ph.D.
Scientific Data Group
Computer Science and Mathematics Division
Oak Ridge National Laboratory
(865) 574-3137 http://www.csm.ornl.gov/~ost
More information about the R-sig-hpc