unusual results with Rmpi
I don't think
mpi.bcast.Robj2slave(sim.dat$bsiaic.dat)
will work because R cannot name sim.dat$bsiaic.dat as an object on slaves. I don't know what exactly your functions do on slaves. Be aware that all slaves are likely to have the same random seed unless parallel rng is used. Install rlecuyer and use mpi.setup.rngstream() to initiate parallel rng. Hao
Maas James Dr (MED) wrote:
I'm running a small programme as a test and to learn how to use Rmpi. I send a data set and routine to two slave nodes, and then get the answers back in an object. The routines executed are Bayesian Gibbs sampling routines so I would expect slightly different answers back from the two nodes but I get identical answers back from both nodes. I'm under the impression that each node does the complete routine from start to finish before returning answers but this may be my misunderstanding. Should the answers in fact be identical because mpi has split the job up between the two nodes and both contributed to the calculations, therefore both end up with the same answer? mpi code that gives me these results is : ## send the data to slaves mpi.bcast.Robj2slave(sim.dat$bsiaic.dat) mpi.bcast.Robj2slave(sim.dat$bsdaic.dat) ## send the jags fitting functions to slaves mpi.bcast.Robj2slave(bsiaicresults) mpi.bcast.Robj2slave(bsdaicresults) ## execute the fitting functions on slaves and ## get results back bsiaictot <- mpi.remote.exec(bsiaicresults()) bsdaictot <- mpi.remote.exec(bsdaicresults()) Any suggestions? Thanks J =============================== Dr. Jim Maas University of East Anglia [[alternative HTML version deleted]]
_______________________________________________ R-sig-hpc mailing list R-sig-hpc at r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-hpc
Department of Statistics & Actuarial Sciences Fax Phone#:(519)-661-3813 The University of Western Ontario Office Phone#:(519)-661-3622 London, Ontario N6A 5B7 http://www.stats.uwo.ca/faculty/yu