[Trilinos-Users] Finite Element Assembly

Danielle Catherine Maddix dcmaddix at stanford.edu
Fri Jan 29 17:17:47 EST 2016


I am getting confusing timing results based on two versions of our finite element code that uses Trilinos.  I started with an in-house finite element solver and we wanted to see how Trilinos linear solvers compared to ours. I am using Epetra and AztecOO.  I have two methods.  One where I copy our global stiffness and global force vector into Trilinos, using BeginReplaceIntoGlobalValues.  My matrix is FeVBRMatrix.  In this case, the times of the linear solver per each Newton iteration are given below:

time_global=[0.1058    0.7401    0.8486    0.8326    0.8607]

iter_global =  [50 196 214 212 218]

In the other case, I replace our current assemble routine and assemble in Epetra using BeginSumIntoGlobalValues.  I verified that the global stiffness matrix and force vector that I get is exactly the same, before I call gmres on it. However, the times are sufficiently slower for the same number of iterations and same solution.

time_assemble = [0.320155 1.6002 1.75062 1.71539 1.84786]

iter_assemble = [50 196 214 212 218]

So I see a 2-3x slowdown for the same number of iterations and same linear problem between the two methods.  Is there any potential reason for this that may come to mind?


Danielle Maddix
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://trilinos.org/pipermail/trilinos-users/attachments/20160129/648a72cb/attachment.html>

More information about the Trilinos-Users mailing list