[R-sig-hpc] Rmpi over mpich2 working, new questions

[Martin Morgan]

Thomas Hampton wrote:
> We worked through our problems and got Rmpi to work on Red Hat, even
> calling it from
> papply. Thanks for your help.
> 
> Interestingly, a  parallelization of 1,000 trivial operations involving
> 7 slaves took  32 seconds on our system. Is it possible that the
> message passing overhead is borne in serial and takes 32
> milliseconds per task?

I don't know how papply does things in particular, but communication
costs can be considerable, especially when trying out 'trivial'
computations. Try realistic tasks (i.e., substantial computation) where
you've thought carefully about minimizing the data to be sent to and
returned from the workers. Personally, I use Rmpi directly (e.g.,
mpi.parLapply) for its flexibility.

Martin

> 
> 
> Thanks again,
> 
> Tom
> 
> _______________________________________________
> R-sig-hpc mailing list
> R-sig-hpc at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-hpc