This is a read only copy of the old FEniCS QA forum. Please visit the new QA forum to ask questions

Unable to project onto large functionspace in newer versions

+1 vote

Hi all!

When using 2016.1.0 or 2016.2.0 I am unable to project to a large functionspace due to "UMFPACK V5.6.1 (Jun 20, 2012): ERROR: out of memory".

A very simple example would be:


from dolfin import *
mesh = UnitCubeMesh(50,50,50)
V = FunctionSpace(mesh,'CG',1)
f = project(Constant(1.0),V)

This does work in 1.6.0 I have running. 2016.2.0 was installed using the Automated build script and then sourcing fenics.stable.

Thanks in advance

asked Dec 23, 2016 by meron FEniCS User (2,340 points)

I can reproduce the error in 2016.1.0 and 2016.2.0 using the Docker images quay.io/fenicsproject/stable:2016.1.0 and quay.io/fenicsproject/stable:latest, while it works fine in 1.6.0 using quay.io/fenicsproject/stable:1.6.0. Please send an email about this to the FEniCS support mailing list.

2 Answers

+2 votes

Have you tried running in parallel? I think UMFPACK is the default solver when running in serial which has some restrictions as you've noticed here with memory.

Try for example using a direct solver:

f = project(Constant(1.0),V, solver_type="mumps")
f = project(Constant(1.0),V, solver_type="superlu")

or a Krylov subspace solver

f = project(Constant(1.0),V, solver_type="cg", preconditioner_type="amg")
answered Dec 23, 2016 by nate FEniCS Expert (17,050 points)
+1 vote

Docker is allocated some memory. You could try increasing this. E.g., I have 16 GB RAM on my laptop and by default docker is getting 2 GB RAM. This can be increased in docker preferences and restarting it.

answered Dec 24, 2016 by praveen FEniCS User (2,760 points)
...