[R] help with random numbers and Rmpi

[Na Li]

On 1 Dec 2003, Faheem Mitha spake thusly:
  
>  I am trying to get a simple program working using Rmpi with the model of 1
>  R master and n C slaves. What I am trying to do is have each of the C
>  slaves generate a random number from U[0,1], and then have the master
>  collect all n numbers as a vector and output it. However even doing this
>  is rather over my head. I'm trying to use rsprng and sprng, but I am sure
>  what I am currently doing is wrong.
>  
>  *************************************************************************
>  rand.R
>  *************************************************************************
>  library(Rmpi)
>  
>  rand <- function ()
>  {
>  if (mpi.comm.size(1) > 1)
>  stop ("It seems some slaves running on comm 1.")
>  mpi.comm.spawn("./rand")
>  mpi.intercomm.merge(2,0,1)
>  
>X  mpi.init.sprng()
>X  free.sprng() #does this function exist here?

I'm not familiar with Rmpi code.  But here you don't need call any SPRNG
function.  Instead, generate a seed (an integer) can pass it on to the
slaves. 
  
>  rdata <- double(sum(mpi.comm.size(1)))
>  out <- mpi.gather(0, 2, rdata) ##this isn't right
>  mpi.comm.free()
>  out
>  }
>  
>  *************************************************************************
>  rand.c
>  *************************************************************************
>  #include <mpi.h>
>  #include <sprng.h>
>  
>  int main(int argc, char **argv)
>  {
>  double rand;
>  double* randarray;
>  
>  MPI_Comm slavecomm, all_processes;
>  
>  /*Initialize MPI*/
>  MPI_Init(&argc, &argv);
>  
>  MPI_Comm_get_parent(&slavecomm);
>  MPI_Intercomm_merge(slavecomm, 1, &all_processes);
>  
>  /*How many processes are there?*/
>  MPI_Comm_size(all_processes, &size);
>  
>  /*Which one am I?*/
>  MPI_Comm_rank(all_processes, &rank);
>  
>X  init_sprng()
>X  rand = sprng();
>X  free.sprng()

Receive the seed from the master and call

   int * stream_id;

   stream_id = init_sprng (gtype, rank, size, seed, param);

where gtype and param can be predefined or got from the master as well.

Now you can generate random numbers by:

   rand = sprng (stream_id);
   free_sprng (stream_id);

>  randarray = (double *)malloc(sizeof(double)*size);
>  
>  /*Gather random numbers from all C slave processes*/
>  /* Using randarray doesn't make sense since this should correspond
>  to the rdata vector in the R master process */
>  MPI_GAther(&rand, 1, 1, MPI_DOUBLE, randarray, 1, MPI_DOUBLE, 0,
>  all_processes);
>  
>  /*All done*/
>  MPI_Comm_free(&all_processes);
>  MPI_Finalize();
>  exit(0);
>  }

There is nothing mysterious about the MPI code in SPRNG, don't use it.
Generating and passing seeds around yourself.

Michael

-- 
Na (Michael) Li, Ph.D.
Assistant Professor
Division of Biostatistics, University of Minnesota
A443 Mayo Bldg, MMC 303         Phone: (612) 626-4765
420 Delaware St SE              Fax:   (612) 626-0660