R help - Dec 2007 - Multicore computation in Windows network: How to set up Rmpi

R-users,

My question is related to earlier posts about benefits of quadcore over
dualcore computers; I am trying to setup a cluster of windows xp
computers so that eventually I could make use of 10-20 cpu:s, but for
learning how to do this, I am playing around with two laptops.

I thought that the package snow would come handy in this situation, but
to use snow, I would probably need to install package Rmpi first. I
might also like to use Rmpi directly.

In order to use Rmpi in windows, I need to install MPI middle-ware, for
which I have found two options:

MPICH2: http://www.mcs.anl.gov/research/projects/mpich2/

DeinoMPI: http://mpi.deino.net/

First I tried MPICH2 1.06 + R-2.6.0 + Rmpi 0.5-5. (I downloaded windows
binaries of Rmpi from Rmpi website
:http://www.stats.uwo.ca/faculty/yu/Rmpi/)

With MPICH2 I managed to connect my computers so that I was able to
remotely launch Rgui on both machines but R hanged when calling
"library(Rmpi)". If only one Rgui was launched on the localhost,
"library(Rmpi)" worked without errors, but trying to use
"mpi.spawn.Rslaves()" resulted in an error message, and so did
"mpi.universe.size()". (In my current setup I can not reproduce this 
error message, but I can go back to this setup if this seems to be an 
important piece of information)

After that I removed MPICH2 from the system and installed DeinoMPI
instead, thus my setup was DeinoMPI 1.1.0 + R-2.6.0 + Rmpi 0.5-6.
Using this setup on a single machine seems to work,
"mpi.universe.size()" returns the correct number of cpu cores and
"mpi.spawn.Rslaves()" creates one master and one slave on the local 
unicore computer. However, trying to use two computers results in 
similar behavior as with MPICH2: Rgui gets started on both, but R hangs 
when trying "library(Rmpi)".

Following the advice given at the Rmpi website, this behavior could be
due to firewall settings. However, the result is the same if I take down
all my firewalls. Trying to debug this, I tried to run the example MPI
programs provided with the DeinoMPI installation. The example programs
work as expected: both machines participate to parallel computation and
the result is shown on the master node. This makes me believe that the
problem is likely related to R and Rmpi configuration, or the settings
I used when launching Rgui using mpiexec:

Settings on Mpiexec- tab:
Application: "C:\Program Files\R\R-2.6.0\bin\Rgui.exe"
Number of processes: 2
hosts: "akva26 samu" These are the computers with DeinoMPI installed.
DeinoMPI cluster-tab shows that they are ready to accept MPI jobs.)
localroot: checked

Other options were just empty.

I am sure that someone has tried these before, and I was hoping to find 
such users from this mailing list. Could you kindly share your 
experiences about this issue?

For example, does anyone have a working setup with DeinoMPI? According 
to Rmpi website DeinoMPI is the easiest way to set up MPI for a single 
windows machine, but I am not sure whether this is also intended to mean 
that one can not expect it to work with multiple computers.

Regards,

Samu M?ntyniemi


------------------------------------------
Samu M?ntyniemi
Researcher
Fisheries and Environmental Management Group (FEM)
Department of Biological and Environmental Sciences
Biocenter 3, room 4414
Viikinkaari 1
P.O. Box 65
FIN-00014 University of Helsinki

Phone: +358 9 191 58710
Fax: +358 9 191 58257

email: samu.mantyniemi  helsinki.fi
personal webpage: http://www.helsinki.fi/people/samu.mantyniemi/
FEM webpage: http://www.helsinki.fi/science/fem/

Some progress in my problem:

Samu M?ntyniemi kirjoitti:
> With MPICH2 I managed to connect my computers so that I was able to
> remotely launch Rgui on both machines but R hanged when calling
> "library(Rmpi)". If only one Rgui was launched on the localhost,
> "library(Rmpi)" worked without errors, but trying to use
> "mpi.spawn.Rslaves()" resulted in an error message, and so did
> "mpi.universe.size()". (In my current setup I can not reproduce
this
> error message, but I can go back to this setup if this seems to be an 
> important piece of information)
I vent back to MPICH2 installation to see what the error was:
"ERROR in names(HOSTNAMES)<-base: attempt to set an attribute on
NULL"

Trying to rethink what the problem was I realized that unlike in 
DeinoMPI, I need to write the host names manually on the "configurable 
settings" -window, and in order to have one cpu available on the local 
machine, I need to write "myhostname:2".

After these changes MPICH2 1.06 +R-2.6.0+Rmpi 0.5-5 work on the single 
machine in the same way as my DeinoMPI installation: Correct number of 
cpu:s is detected and I can "mpi.spawn.Rslaves()"

I will try to do this with two hosts next and see if there is more luck 
with MPICH2 than DeinoMPI.

Samu




------------------------------------------
Samu M?ntyniemi
Researcher
Fisheries and Environmental Management Group (FEM)
Department of Biological and Environmental Sciences
Biocenter 3, room 4414
Viikinkaari 1
P.O. Box 65
FIN-00014 University of Helsinki

Phone: +358 9 191 58710
Fax: +358 9 191 58257

email: samu.mantyniemi at helsinki.fi
personal webpage: http://www.helsinki.fi/people/samu.mantyniemi/
FEM webpage: http://www.helsinki.fi/science/fem/

Hello!

I finally got MPICH 1.06 + R 2.6.1 + Rmpi 0.5-5 working with multiple 
computers. The key was to realize that the number of processes should be 
one when launching Rgui using mpiexec and not the number of 
master+slaves, as I had first wrongly understood.

However, I seem to have a new problem which I have not been able to 
figure out:

After loading Rmpi, the first attempt to mpi.spawn.Rslaves() always 
spawns the slaves on the local machine instead of on both machines. If I 
close the slaves and spawn again, then one slave gets spawned on remote 
machine. Each time I close and then spawn againg, the order of machines 
is different, and eventually I get back to the situation where all 
slaves are on the local machine. Continuing to do spawning and closing 
seems to reveal a pattern. I can see similar behavior if I have more 
than two machines, and it takes more spawn-close cycles to get all my 
slave machines spawned on.

Below is an example session with two machines. This pattern shows 
everytime I start R and run this script. How to control the spawning so 
that I get everything right at the first call of mpi.spawn.Rslaves()?

Regards,

Samu

<R>

 >
 > library(Rmpi)
 > sessionInfo()
R version 2.6.1 (2007-11-26)
i386-pc-mingw32

locale:
LC_COLLATE=Finnish_Finland.1252;LC_CTYPE=Finnish_Finland.1252;LC_MONETARY=Finnish_Finland.1252;LC_NUMERIC=C;LC_TIME=Finnish_Finland.1252

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base

other attached packages:
[1] Rmpi_0.5-5
 > mpi.universe.size()
[1] 2
 > mpichhosts()
          master          slave1          slave2
"clustermaster" "clustermaster" "clusterslave1"
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterMaster
slave2 (rank 2, comm 1) of size 3 is running on: ClusterMaster
 > mpi.close.Rslaves()
[1] 1
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterSlave1
slave2 (rank 2, comm 1) of size 3 is running on: ClusterMaster
 > mpi.close.Rslaves()
[1] 1
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterMaster
slave2 (rank 2, comm 1) of size 3 is running on: ClusterSlave1
 > mpi.close.Rslaves()
[1] 1
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterMaster
slave2 (rank 2, comm 1) of size 3 is running on: ClusterMaster
 > mpi.close.Rslaves()
[1] 1
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterSlave1
slave2 (rank 2, comm 1) of size 3 is running on: ClusterMaster
 > mpi.close.Rslaves()
[1] 1
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterMaster
slave2 (rank 2, comm 1) of size 3 is running on: ClusterSlave1
 > mpi.close.Rslaves()
[1] 1
 >
 >
 > mpi.spawn.Rslaves()
         2 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 3 is running on: ClusterMaster
slave1 (rank 1, comm 1) of size 3 is running on: ClusterMaster
slave2 (rank 2, comm 1) of size 3 is running on: ClusterMaster
 > mpi.close.Rslaves()
[1] 1
 >

</R>


Samu M?ntyniemi kirjoitti:
> Some progress in my problem:
> 
> Samu M?ntyniemi kirjoitti:
> 
>> With MPICH2 I managed to connect my computers so that I was able to
>> remotely launch Rgui on both machines but R hanged when calling
>> "library(Rmpi)". If only one Rgui was launched on the
localhost,
>> "library(Rmpi)" worked without errors, but trying to use
>> "mpi.spawn.Rslaves()" resulted in an error message, and so
did
>> "mpi.universe.size()". (In my current setup I can not
reproduce this
>> error message, but I can go back to this setup if this seems to be an 
>> important piece of information)
> 
> I vent back to MPICH2 installation to see what the error was:
> "ERROR in names(HOSTNAMES)<-base: attempt to set an attribute on
NULL"
> 
> Trying to rethink what the problem was I realized that unlike in 
> DeinoMPI, I need to write the host names manually on the "configurable
> settings" -window, and in order to have one cpu available on the local
> machine, I need to write "myhostname:2".
> 
> After these changes MPICH2 1.06 +R-2.6.0+Rmpi 0.5-5 work on the single 
> machine in the same way as my DeinoMPI installation: Correct number of 
> cpu:s is detected and I can "mpi.spawn.Rslaves()"
> 
> I will try to do this with two hosts next and see if there is more luck 
> with MPICH2 than DeinoMPI.
> 
> Samu
> 
> 
> 
> 
> ------------------------------------------
> Samu M?ntyniemi
> Researcher
> Fisheries and Environmental Management Group (FEM)
> Department of Biological and Environmental Sciences
> Biocenter 3, room 4414
> Viikinkaari 1
> P.O. Box 65
> FIN-00014 University of Helsinki
> 
> Phone: +358 9 191 58710
> Fax: +358 9 191 58257
> 
> email: samu.mantyniemi at helsinki.fi
> personal webpage: http://www.helsinki.fi/people/samu.mantyniemi/
> FEM webpage: http://www.helsinki.fi/science/fem/
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 
> 

-- 
------------------------------------------
Samu M?ntyniemi
Researcher
Fisheries and Environmental Management Group (FEM)
Department of Biological and Environmental Sciences
Biocenter 3, room 4414
Viikinkaari 1
P.O. Box 65
FIN-00014 University of Helsinki

Phone: +358 9 191 58710
Fax: +358 9 191 58257

email: samu.mantyniemi at helsinki.fi
personal webpage: http://www.helsinki.fi/people/samu.mantyniemi/
FEM webpage: http://www.helsinki.fi/science/fem/