Running FEniCS on OS X vs Ubuntu


207
views
0
3 months ago by
Hello,

I'm running simulations on some code that is similar to the code in https://github.com/hplgit/fenics-tutorial/blob/master/pub/python/vol1/ft09_reaction_system.py , the main difference being that the mesh in my code is much larger (2.2 million cell 3D mesh).

I'm using the GMRES solver with ILU preconditioner to solve the linear system (9.4 million x 9.4 million matrix). On OS X, installed via anaconda, it takes about 2 min per iteration, but the same code on Ubuntu takes 40 minutes per iteration. I get a message in the terminal that says 'PETSc Krylov solver starting to solve 9442208 x 9442208 system' and it moves to the next iteration after 40 minutes.

I'm using similar hardware - the Mac has an i5 processor with 8 GB RAM, and Ubuntu is running on a Dell with an i7 processor with 8 GB RAM. I'm able to run simulations on a model mesh that is smaller (1million x 1million linear system) on Ubuntu with no problems.

Has anyone else had a similar experience, or knows what the problem might be? Any help fixing this will be greatly appreciated.

Thank you,
Community: FEniCS Project
The answer depends on how you installed the software.

Also, I am curious to know how the solution compares when docker is used.

Aside, it is common to refer to the size of the problem being approximated by "degrees of freedom" \(n\) rather than by e.g., "with an \(n \times n \) matrix.". This is because the system matrices for discrete DEq's are square.  The reason PETSc prints the matrix size is that it supports solving rectangular matix problems.
written 3 months ago by pf4d  
Thank you for replying. I have installed FEniCS on Ubuntu via the following commands:

sudo add-apt-repository ppa:fenics-packages/fenics
sudo apt-get update
sudo apt-get install --no-install-recommends fenics
sudo apt-get dist-upgrade

=====

I installed FEniCS on Mac as a conda package. I don't have Docker installed. 

For now, I'm running the simulations on the Mac, but am very surprised by how much longer it takes PETSc to solve the 9.4 million dof system on Ubuntu than it does on the Mac with similar hardware specs. I tried reinstalling FEniCS on Ubuntu too.

Any suggestions to speed up the solution on Ubuntu?

Thank you,
written 3 months ago by Sophia Wright  
I suspect that the cobra packages have been optimized, but I do not expect the performance difference to be that drastic... someone else may know.
written 3 months ago by pf4d  
I can say that I have only ever used fenics on Ubuntu and it has performed marvelously for me throughout the years.
written 3 months ago by pf4d  
I've had no problems with FEniCS on Ubuntu until I worked with a very large mesh. Thank you for taking the time to reply - I hope I hear from someone who has experienced a similar problem,
written 3 months ago by Sophia Wright  
I have solved problems much larger than this using Ubuntu using the repo install method, and have not experienced the problem. Perhaps you can verify that the solvers are identical?
written 3 months ago by pf4d  
Just verified that the code is exactly the same - same solver too - GMRES with ILU preconditioner. Is there any other way to verify that the solvers are identical?
written 3 months ago by Sophia Wright  
Please login to add an answer/comment or follow this question.

Similar posts:
Search »