svsolver + mpiexec bug

[Julian Suk]

Hi!

I am running svsolver with mpiexec from a terminal in Ubuntu 20.04 on two separate machines (lab-wide server and personal workstation). It ran flawlessly on both machines about a month ago.

When I try to run svsolver with mpiexec on my workstation now, I get an error message "Invalid MIT-MAGIC-COOKIE-1 key" for each process I try to spawn. The folder "8-procs_case" which is usually used to run the parallelisation is not created. Then each of the parallel processes seems to solve the same system, i.e. the computation is not properly parallelised.

I checked if the same code would properly run on the server which it did: the folder "8-procs_case" was created and the computation was parallelised. The server probably did not receive system-wide updates since November unlike my workstation which I update regularly.

I am aware that this might be rather an OpenMPI problem, but updating openmpi-bin did not help and I am clueless how to solve this.

Thanks in advance,

Julian

[Julian Suk]

I figured out what went wrong and am sharing my fix for anyone that might run into similar problems. It appears that a recent update in my Anaconda installation has installed mpiexec and made this version the default when operating from within the Anaconda environment.

Code: Select all

(base) $ which mpiexec
/home/julian/anaconda3/bin/mpiexec
(base) $ readlink -f "/home/julian/anaconda3/bin/mpiexec"
/home/julian/anaconda3/bin/mpiexec.hydra

The latest svSolver however, wants OpenMPI (or its synonym "orterun"):

Code: Select all

$ readlink -f "/usr/local/sv/svsolver/2021-09-30/bin/mpiexec"
/usr/bin/orterun

So, running svSolver via

Code: Select all

(base) $ /usr/bin/orterun -np 8 /usr/local/sv/svsolver/2021-09-30/svsolver

solved the problem.

[David Parker]

Hi Julian,

Thanks for sharing this fix!

Cheers,
Dave