Distributed computing in R with Rmpi

From ScientificComputing
Revision as of 19:52, 10 December 2020 by Jarunanp (talk | contribs)

Jump to: navigation, search

Load modules

Change to the new software stack and load required modules. Here we need MPI and R libraries.

$ env2lmod
$ module load gcc/6.3.0 openmpi/4.0.2 r/4.0.2

Run R in an interactive session

Rmpi assigns one processor to be the master and other processors to be workers. Here we would like to use 4 processors for computation. Therefore, we request 5 processors

 $ bsub -n 5 -W 02:00 -I bash
 Generic job.
 Job <155200980> is submitted to queue <normal.4h>.
 <<Waiting for dispatch ...>>
 <<Starting on eu-c7-105-05>>

Define available global number of processors with the environment parameter MPI_UNIVERSE_SIZE.


Start R

 $ R

Use Rmpi

1. Load Rmpi which calls mpi.initialize()

 > library(Rmpi)

2. Spawn R-slaves to the host. nslaves = requested number of processors - 1

 > usize <- as.numeric(Sys.getenv("MPI_UNIVERSE_SIZE"))
 > ns <- usize - 1
 > mpi.spawn.Rslaves(nslaves=ns)

3. Set up a variable array

 > var = c(11.0, 22.0, 33.0)

4. Root sends state variables and parameters to other ranks

 > mpi.bcast.data2slave(var, comm = 1, buffunit = 100)

5. Get the rank number of that processor

 > mpi.bcast.cmd(id <- mpi.comm.rank())

6. Check if each rank can use its own value

 > mpi.remote.exec(paste("The variable on rank ",id," is ", var[id]))

7. Root orders other ranks to calculate

 > mpi.bcast.cmd(output <- var[id]*2)

8. Root orders other ranks to gather the output

 > mpi.bcast.cmd(mpi.gather(output, 2, double(1)))

9. Root gathers the output from other ranks

 > mpi.gather(double(1), 2, double(usize))

10. Close down and quit

 > mpi.close.Rslaves(dellog = FALSE)
 > mpi.quit()


  1. What happens when replacing mpi.scatter.Robj() instead of mpi.bcast.data2slave() in point 4?
  2. Create an R script using Rmpi and submit a batch job through BSUB command line
  3. Create a BSUB job script and submit a batch job

Further reading