parallel library's random numbers

Ross,

Consider writing this SPMD style and run as Rscript. Here is a short
example:

library(pbdMPI, quiet=TRUE)
init()
comm.set.seed(1234567, diff=TRUE)
r <- runif(5)
comm.print(r, all.rank=TRUE)
finalize()

Put the above 6 lines of R code into file rand.r and then run
mpirun -np 2 Rscript rand.r
Experiment with diff=FALSE and with more than 2 nodes. Then start
reading the vignette for the pbdDEMOpackage.

If diff=TRUE, the streams are independent. This uses the package rlecuyer.

Good luck!

George

P.S. I assume your 50 nodes are in a cluster, probably managed with PBS.
Then the mpirun is issued after a "qsub -I ..." interactive allocation.

On 12/19/13 5:08 PM, Ross Boylan wrote:

On Thu, 2013-12-19 at 13:28 -0800, Ross Boylan wrote:

Conceptually, I have 2,000 tasks and I want each task to have access to
an independent, repeatable random number stream.  Many of the tasks will
run on the same node.  For example, a typical job would be to do tasks
200 to 400.

I might do one run with 10 nodes and a later one with 50, but I want the
same streams for each task.

The facilities in parallel seem designed to achieve repeatable
randomization for a fixed number of nodes.  Is there a way to get them
to do what I want?

I'm thinking maybe I should just ignore the parallel random number stuff
and do set.seed(t) before task t.  It's easy to imagine that is not an
entirely safe method, however.

Ross

The documentation does not explain what nextRNGStream and
nextRNGSubStream do nor how that relates to the initialization done by
clusterSetRNGStream.  For example, if I do nextRNGStream on node 1, do I
get the same stream as is being used on node 2?

The documentation does not say that calling nextRNGStream actually
resets the seed, though that seems to be implicit in the example (and is
explicit in one book I found on the net).  I'm also unsure if calling
mc.reset.stream() is necessary after calling clusterSetRNGStream.

Ross Boylan

P.S. Using R 3.0.1