Ubuntu 12.04 and MPICH performance

Question

I am trying to install mpich-3.1 in a linux cluster (Ubuntu 12.04 running on all machines). Previously I installed mpich2 by sudo apt-get install mpich2 but couldn't find how to run tests. Then I removed with sudo apt-get remove mpich2.

So I decided to upgrade to 3.1. I downloaded and installed mpich following instructions at https://www.mpich.org/static/downloads/3.1/mpich-3.1-installguide.pdf by running:

sudo ./configure -prefix=/usr/local/mpich/

sudo make

sudo make install

and apparently is properly installed. If I run

meteo@ventus:~/RAMS/RUN$ /usr/local/mpich/bin/mpiexec -f machinefile -n 20 hostname
ventus
ventus
ventus
ventus
ventus4
ventus4
ventus4
ventus4
ventus5
ventus5
ventus5
ventus5
ventus2
ventus2
ventus2
ventus2
ventus3
ventus3
ventus3
ventus3

Although I find it is responding "slowly". Where machinefile is

ventus:4
ventus2:4
ventus3:4
ventus4:4
ventus5:4

The directory is exported to all nodes in the cluster, /etc/exports

/usr/local/mpich 192.168.1.0/24(rw,sync)

In /etc/mtab ant /etc/fstab in node ventus4

ventus:/usr/local/mpich /usr/local/mpich nfs rw,vers=4,addr=192.168.1.1,clientaddr=192.168.1.4 0 0

ventus:/usr/local/mpich /usr/local/mpich   nfs

Maybe the problem comes from a prior install not completely removed

meteo@ventus:~$ which mpiexec
/usr/local/bin/mpiexec
meteo@ventus:~$ which mpirun
/usr/local/bin/mpirun
meteo@ventus:~$ which mpicc
/usr/local/bin/mpicc

Following installation instructions which mpiexec should point to mpich installation bin directory /usr/local/mpich/bin/mpiexec

But if I move /usr/local/bin/mpiexec to /usr/local/bin/mpiexec.old

then

meteo@ventus:~$ which mpiexec
/usr/local/mpich/bin/mpiexec

points to my new mpich3 install directory. Could this be the reason of that slow performance? Which test should I run for benchmarking? How do I completely remove mpich2.

Answer 1

If you do in fact have root access to your entire machine, then you can always just delete all of the binaries, libraries, headers, etc. I'm not sure where everything is installed on your system (it's different everywhere), but the usual locations are /usr/local/bin, /usr/local/include, /usr/local/lib, etc. You should look for these files (or things that look similar:

bin/:

• Anything that starts with mpi

include/:

• Anything that starts with mpi
• Anything that starts with opa

lib/:

• Anything that includes mpich
• Anything that includes mpl
• Anything that includes opa

Beyond that, there's not much that would interfere (there's man pages somewhere too, but that's fine). If you delete all of those files, you should have gotten rid of your MPICH2 installation. This really should have been cleaned up with you did your apt-get uninstall, but that's neither here nor there...

Now, to test your new MPICH installation (the project is called MPICH now, not MPICH3), there are lots of MPI benchmarks. I'd suggest typing mpi benchmarks into your favorite search engine any trying out a few of the ones that you find. If you want to compare, you can install a few different versions of MPI. When you do this, make sure you are correctly setting up your PATH and LD_LIBRARY_PATH environment variables so your installations can sit side by side.