[R-sig-hpc] unstable repeated spawning with Rmpi under Windows 7
hyu at stats.uwo.ca
Thu Dec 13 16:48:16 CET 2012
Windows always behaves different, in particular memory allocation and
release. I don't think repeated spawning is good even for MPI itself. Why
don't you keep the session and continue your computation?
One possible solution is to add a sleep period before next spawning. Try
(sleep 5 sec) before next spawning.
Till Francke wrote:
> Dear list,
> I observed irregular hang-ups when repeatedly performing this code (R
> 2.15.0, Windows 7 64, MPICH-II, Rmpi_0.6-2):
> When I repeatedly call this code, my system hangs at
> "mpi.spawn.Rslaves(nslaves=8)" - sometimes at the second, sometimes at the
> third try. Then, killing R remains the only option.
> When I use the program "Process Explorer" to see what is happening I
> observed that mpi.spawn.Rslaves() creates mpiexec.exe and a service
> smpd.exe, comprising a tree of 8 consoles (cmd.exe) each holding an
> instance of R (Rterm.exe). Apparently, mpi.close.Rslaves() is supposed to
> close these. However, evidently this sometimes fails and the subsequent
> call of mpi.spawn.Rslaves() stalls the system. If I remove the consoles
> (but not the parent smpd.exe) manually, the next call works as supposed.
> Is there anything I can do about it (except switching to Linux)?
> Thanks for your advice,
> R version 2.15.0 (2012-03-30)
> Platform: x86_64-pc-mingw32/x64 (64-bit)
>  LC_COLLATE=German_Germany.1252 LC_CTYPE=German_Germany.1252
>  LC_NUMERIC=C LC_TIME=German_Germany.1252
> attached base packages:
>  stats graphics grDevices utils datasets methods base
> other attached packages:
>  Rmpi_0.6-2
> R-sig-hpc mailing list
> R-sig-hpc at r-project.org
Department of Statistics & Actuarial Sciences
The University of Western Ontario
London, Ontario N6A 5B7
More information about the R-sig-hpc