MUMPS workspace too small

Clearly defined bug reports and their fixes
Post Reply
fgillet
Posts: 47
Joined: 30 Sep 2010, 16:58

MUMPS workspace too small

Post by fgillet »

Dear Elmer Developpers;

I've had troubles recently with mumps for big parallel simulations;
The error message from mumps is
WORKSPACE TOO SMALL FOR DMUMPS_
ERROR RETURN ** FROM DMUMPS INFO(1)= -9
etc...

It's possible to solve this problem by defining mumps control parameter icntl(14) (i.e by adding A % MumpsID % icntl(14) = ... in DirectSolve.f90) to change the default value of 20.
I think the best, would be to add a SolverKeyword, so that we can change this value from the sif file in case of troubles;

Best regards

Fabien
Juha
Site Admin
Posts: 357
Joined: 21 Aug 2009, 15:11

Re: MUMPS workspace too small

Post by Juha »

Hi Fabian,

>I think the best, would be to add a SolverKeyword, so that we can change this value from the sif file in case of troubles;

i'll see what i can do. Thanks for the report!

Juha
pamakela
Posts: 49
Joined: 06 Mar 2010, 18:37

Re: MUMPS workspace too small

Post by pamakela »

Hi,


I get this error as well. I tried compiling after adding "A % MumpsID % icntl(14) = ..." in DirectSolve.f90 and DirectSolve.src but it didn't solve the problem.

This error occurred while running elmersolver_mpi with a GMSH generated mesh with 2nd order elements. (will still try with 1st order elements and see what happens)

I wonder what the workflow is to fix this?
lionel
Posts: 27
Joined: 26 Jan 2010, 12:32

Re: MUMPS workspace too small

Post by lionel »

Hi Elmer,

I've had the same troubles lately, happening when I share my simulation among more than 2 nodes. Sometimes there is the same error message than Fabien. Sometimes there is nothing but a Seg Fault whereas the same simulation works with a low number of partitions.

Best Regards
Lionel
pamakela
Posts: 49
Joined: 06 Mar 2010, 18:37

Re: MUMPS workspace too small

Post by pamakela »

Hi,


I was able to alleviate much of these 'workspace too small' errors by introducing preconditioning (HYPRE) and running my simulation with smaller timesteps. Eventually I still get the error but I think that in my case I can produce enough simulation data before the error halts the run.
tzwinger
Site Admin
Posts: 100
Joined: 24 Aug 2009, 12:20
Antispam: Yes

Re: MUMPS workspace too small

Post by tzwinger »

I know that this reply is vastly delayed, but I think I insert this piece of information (beside the fact that it has been mentioned in the latest release notes) also here for those who are looking up this thread: The code lines Fabien suggested have been inserted and the corresponding MUMPS parameter inctl(14) can be set now to another value than its default of 20, for instance 60, by the line:

Code: Select all

Mumps percentage increase working space = Integer 60
This has to be inserted into the specific Solver section of the SIF.

Best wishes

Thomas
Post Reply