ValueError: Too large work array required -- computation cannot be performed with standard 32-bit LAPACK

[MaartenDie]

When I used pyamg's multigrid ruge_stuben_solver I get the mentioned error. The strange thing, for one matrix A the solver works. However, when I alter the matrix along the middle diagonal I get the error. Is this a known issue?

Thanks!

Maarten

[lukeolson]

Not a known issue as far as I know. To help track it down, do you have a small example?

Does pyamg.test() pass?

Also, some package info would help:

import sys, numpy, scipy, pyamg
print(sys.version, numpy.__version__, scipy.__version__,  pyamg.__version__, sep='\n')

[MaartenDie]

Thanks for the quick reply!

Unfortunately it's not possible to provide a small example. When I calculate the same problem with a discretisation resulting in a matrix of 6000x6000 the solver works. When I increase the discretisation with a factor 2 resulting in a matrix of 12000x12000 I get the former mentioned error. Next week I'm going to try to reduce the code and provide a simple, but unfortunately large, example. In the mean time I did find a work around, but I'm still curious about the error. Thanks for the help.

[bensworth]

How big is your coarsest matrix? In particular are you doing two-level? That is the only place I can think of LAPACK throwing an error in RS.

If not, have you used pdb before? Would be useful if you stepped through the code so we knew more specifically where the error happens. I don't recall anywhere in RS that we should be using LAPACK aside from potentially a dense solve on the coarsest level.

[lukeolson]

That's my guess too. As a start, can you paste the pyamg call that throws the error?

Something like pyamg.ruge_stuben_solver(... and ml.solve(...

[MaartenDie]

The coarsest matrix is a pentadiagonal matrix Poisson matrix with some empty elements. The other diagonals are sparse. The large matrix is 12000x12000, here the error came up. The same problem for 6000x6000 did work. A different problem with a different values on the first diagonal on a matrix of 12000x12000 did work. The final call I made was:

ml = pyamg.ruge_stuben_solver(temp_LHS)
x_new = ml.solve(temp_RHS, tol=1e-6, x0=x_old)

I have only used pdb in my thesis on which I'm working right now. I assume two level means the level on which the V-cycle happens. I do no specify a level in a call, but I think the error happens on the fifth level. The program I use is called Transient.py, the exact error I get is in the document
slurm-499260.txt

Thanks in advance!

[lukeolson]

Ok, this is helpful. I'll try to reproduce ... it does look like a coarse-level issue. Can you provide

import sys, numpy, scipy, pyamg
print(sys.version, numpy.__version__, scipy.__version__,  pyamg.__version__, sep='\n')

[MaartenDie]

The response is the following:

3.8.12 (default, Jun 16 2022, 10:30:18)
[GCC 8.5.0 20210514 (Red Hat 8.5.0-10)]
1.19.5
1.9.0
4.2.3