Aloft: R Package Rmpi Installation

More...

This page is copied from Debian Documentation.

Install LAM-MPI
Cluster without MPI_Comm_spawn
Install Rmpi
Test
Final note

Rmpi is an R package providing an interface to MPI (Message-Passing Interface) API calls with interactive R slave functionalities.

Rmpi is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA

Rmpi requires R version 2.0.0 or higher, and LAM-MPI 7.0.2 or higher on unix platform or MPICH2 for Windows on windows platform. Instructions to install and run Rmpi on windows platform are given in http://www.stats.uwo.ca/faculty/yu/Rmpi.

Install LAM-MPI#

http://www.lam-mpi.org

Follow the installation instructions provided by LAM-MPI or simply choose a pre-compiled Linux RPM. Modify the file lam-bhost.def (probably in /etc/lam) to include host names and CPU numbers per host in a cluster. A remote shell program such as rsh or ssh must be set properly otherwise LAM-MPI will not be functional. This requires no password prompt when remotely executing a program. Check LAM-MPI documents or FAQ to configure the default remote shell program. Notice that LAM-MPI must be installed on all hosts within a cluster. Make sure LAM executable programs are in the PATH.

On Debian system, it is sufficient to install

lam3 lam3-dev lam-runtime

Once LAM-MPI has been installed and properly configured, login as a non-root user and boot to LAM by

lamboot -v

and run LAM test suite to see if it works properly. Try to run

lamexec C hostname

to see all nodes response with their host names.

NOTE: If your site has a different MPI such as MPICH installed, Rmpi could still detect it and disable MPI-2 related functions automatically. Unless you remove the part of checking lam-mpi in zzz.R.in file in R directory, you need to create a file lamnodes and put the line

echo "FAKE lamnodes"

Make sure lamnodes is executable and in the path. Try

mpirun -np 1 R

to start a master R and load Rmpi.

Cluster without MPI_Comm_spawn#

If your cluster does not have MPI_Comm_spawn enabled so you cannot use mpi.spawn.Rslaves, there are two ways to launch multiple Rs.

a) Save Rprofile in inst (will be in the root directory of Rmpi after installation) as .Rprofile in the working or root directory. Then run, for example,

mpiexec -n 9 R --no-save -q

This will create one master and 8 slaves. Those slaves will behave as though mpi.spawn.Rslaves has been used.

b) To run multiple Rs without using .Rprofile, add the following line

R_PROFILE=${R_HOME}/library/Rmpi/Rprofile; export R_PROFILE

in the R shell script (after "export R_HOME") and rename it, say, Rmpi. Then one can use

mpiexec -n 9 Rmpi --no-save -q

assuming that Rmpi is in the PATH.

Install Rmpi#

If LAM-MPI is installed in /usr or /usr/local, run

R CMD INSTALL Rmpi_version.tar.gz

On Debian with lam3, lam3-dev lam-runtime installed, above will work too. Rmpi will also detect if MPICH installed in /usr/lib as either /usr/lib/mpi or /usr/lib/mpich.

For LAM-MPI or other MPI in a different location, use

R CMD INSTALL Rmpi_version.tar.gz --configure-args=--with-mpi=/mpipath

Notice that R and Rmpi have to be installed on all hosts in a cluster. The paths (directories) to R and Rmpi must be the same on all hosts in a cluster otherwise Rmpi may have difficult to spawn R slaves. Using a NFS file system and putting R and Rmpi on it should work. The same issue applies to user's home directory: it should be the same across a cluster of computers.

Test#

Before running R and Rmpi, make sure LAM-MPI is running properly. Use

lamboot -v

to boot to LAM and use

lamclean

to clean up all MPI states. All LAM-MPI related tasks should be not be run from root. After finishing a MPI related job, use

lamhalt

to shut down LAM.

Make sure R is in the PATH for whole cluster. Try

rsh a_remote_host_machine R --slave

run a simple command such as 3+8 to see it works. If R is not in PATH or remote machines have difficult find it, one solution is to modify the Rslave.sh file in inst directory and put an absolute path to R.

Load Rmpi library in R (assuming successful) and spawn Rslaves by

mpi.spawn.Rslaves(nslaves=1)  #spawn only one slave
mpi.close.Rslaves()           #close slaves
mpi.spawn.Rslaves()           #spawn slaves to use all CPUs

Execute some simple commands to R slaves, e.g.,

mpi.remote.exec(mpi.get.processor.name())
mpi.remote.exec(double(10))
mpi.remote.exec(double(n)) #will not work if n does not exit on slaves
n <- 10
mpi.remote.exec(double, n) #the arg n is passed to double

If package rlecuyer or rsprng has been installed, try to use mpi.setup.rngstream or mpisetup.sprng initialize one (not two).

There are a few demo examples from Rmpi. Use

demo(package="Rmpi")

to find them. Try

demo("simplePI.R")

and run (run mpi.spawn.Rslaves() first)

simple.pi(1000000)

Try other demos to see if Rmpi works properly:

slave1PI.R and slave2PI.R have two functions each: master and slave functions. User should use mpi.spawn.Rslaves to spawn R slaves first and then run the master function with proper arguments. Make sure n/maxjoblen is at least equal to or bigger than total number of slaves. Making it twice as big as the total slaves will show load balancing effect.
simPI.R has a plot in it so trying it in X-windows.
cslavePI.R and masterslave.R use mpi.comm.spawn to spawn their own slaves. Close R slaves if they are alive.

Final note#

It is extremely important that users should close R slaves and Rmpi properly before quiting R. Basically, use mpi.close.Rslaves to close R slaves and use mpi.exit to exit Rmpi environment or use mpi.quit to quit Rmpi and R together.

To safe guard at site level, administrator can add the .Last function to Rprofile.site as

.Last <- function(){
	if (is.loaded("mpi_initialize")){
		if (mpi.comm.size(1) > 0){
			print("Please use mpi.close.Rslaves() to close slaves.")
			mpi.close.Rslaves(comm=1)
		}
		print("Please use mpi.quit() to quit R")
		.Call("mpi_finalize")
	}
}