[R-sig-hpc] Shared Memory

[Markus Schmidberger]

Nathan S. Watson-Haigh schrieb:
> I'm new to HPC and parallel programming but I've created my own R
> package which has a parallel (using Rmpi) and non-parallel
> implementation of the same algorithm. It works nicely, but I'm trying to
> better understand how/if Rmpi uses shared memory. Does/can Rmpi use
> shared memory? For instance, if each slave needs access to the same data
> matrix, does Rmpi create a copy of that data for each slave when I do:
> mpi.bcast.Robj2slave(myMatrix)
>   
Yes
> I think it does create a copy and I therefore currently pass a subset of
> the data matrix to each slave using:
> objList <- list(m=m[xMin:nrow(m), xMin:nrow(m)])
> mpi.send.Robj(objList, slave_id, 1)
>   
Yes, this works
> myMatrix can be up to 24k x 24k in size. That's > 4Gb of RAM just to
> hold it in memory! I suppose I'm wondering if memory requirements are
> proportional to the number of slaves requested - if each slave has to
> have it's own copy of the data and if I can reduce this requirement by
> utilising the shared memory!?
>   
Yes, you will get some memory problems. Using Rmpi I think there is no 
way to use shared memory.
There is a great new package for shared memory sytems: multicore
http://cran.r-project.org/web/packages/multicore/index.html
I already tested it with 500 cores and there is a great performance 
(nearly linear!)

Best
Markus
> Cheers,
> Nath
>
> _______________________________________________
> R-sig-hpc mailing list
> R-sig-hpc at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-hpc
>   


-- 
Dipl.-Tech. Math. Markus Schmidberger

Ludwig-Maximilians-Universität München
IBE - Institut für medizinische Informationsverarbeitung,
Biometrie und Epidemiologie
Marchioninistr. 15, D-81377 Muenchen
URL: http://www.ibe.med.uni-muenchen.de 
Mail: Markus.Schmidberger [at] ibe.med.uni-muenchen.de
Tel: +49 (089) 7095 - 4497