OpenSeesMP with Mumps

This forum is for issues related to parallel processing
and OpenSees using the new interpreters OpenSeesSP and OpenSeesMP

Moderator: selimgunay

Post Reply
brag006
Posts: 173
Joined: Wed Feb 15, 2012 1:26 pm
Location: University of Auckland

OpenSeesMP with Mumps

Post by brag006 » Sun Apr 03, 2016 10:58 pm

Hi,

I am running an analysis using OpenSeesMP on a cluster and using the Mumps solver. I am getting the following error:

WARNING MumpsParallelSolver::solve(void)- Error -9 returned in substitution dmumps()
Work array too small; use -ICNTL14 option, the default is -ICNTL 20 make 20 larger

The interesting thing is I only get that error when I run more the 4 tasks. I have set the system command as follows:

system Mumps -ICNTL14 100

I tried numbers smaller than 100 but with no luck.

I can't seem why it fails only when i run more than 4 tasks in parallel.

fmk
Site Admin
Posts: 5883
Joined: Fri Jun 11, 2004 2:33 pm
Location: UC Berkeley
Contact:

Re: OpenSeesMP with Mumps

Post by fmk » Mon Apr 04, 2016 10:42 am

as you change the number of tasks it may be changing the elimination ordering so that more work can be done in parallel.

it should of course not fail.

brag006
Posts: 173
Joined: Wed Feb 15, 2012 1:26 pm
Location: University of Auckland

Re: OpenSeesMP with Mumps

Post by brag006 » Mon Apr 04, 2016 9:30 pm

Any suggestions on solving this problem? It seems to be a weird and I can't find the source of the problem

fmk
Site Admin
Posts: 5883
Joined: Fri Jun 11, 2004 2:33 pm
Location: UC Berkeley
Contact:

Re: OpenSeesMP with Mumps

Post by fmk » Tue Apr 05, 2016 9:38 am

Mumps is proving to be somehwat tempermental. it does not like for example repeated calls to setSize. sometimes even with the same model one works and the next fails! i am working on the petsc interface in the hope that it is my code hat is messed up and that another interface to Mumps will not have the issues.

brag006
Posts: 173
Joined: Wed Feb 15, 2012 1:26 pm
Location: University of Auckland

Re: OpenSeesMP with Mumps

Post by brag006 » Tue Apr 05, 2016 2:15 pm

Thanks for that, your work is much appreciated.

Do you have any advice on temporary solutions? Time is of the essence and this is a really big obstacle for me.

Any suggestions would be much appreciated.

Regards,

fmk
Site Admin
Posts: 5883
Joined: Fri Jun 11, 2004 2:33 pm
Location: UC Berkeley
Contact:

Re: OpenSeesMP with Mumps

Post by fmk » Thu Apr 07, 2016 9:17 am

other than the obvious (stick with 4 or less for the moment), no .. i am trying but it is just me (and 50% of that at the moment due to funding considerations) and i am juggling a number of hot topic items just now.

Post Reply