Author Topic: Parallel performance of 3D FEM codes  (Read 8676 times)

kashraf

  • New Member
  • *
  • Posts: 3
Parallel performance of 3D FEM codes
« on: June 14, 2013, 02:38:36 PM »
I was wondering what is a typical parallel performance of state of the art FEM code ? Meaning, up to how many processors have linear speed up have been achieved and which codes have achieved it ? What are the bottlenecks in further parallelization?   

Prof. S. Govindjee

  • Administrator
  • FEAP Guru
  • *****
  • Posts: 1164
Re: Parallel performance of 3D FEM codes
« Reply #1 on: June 16, 2013, 10:15:48 AM »
Speed up in parallel codes is a very tricky things.  It depends strongly on the problem you are solving.  No general statements are really possible.  With respect to FEAP, we use Petsc for all the parallel heavy lifting.  If you look at the end of the parallel manual you can see some basic timing results (albeit from about 6 years back).

kashraf

  • New Member
  • *
  • Posts: 3
Re: Parallel performance of 3D FEM codes
« Reply #2 on: June 19, 2013, 12:45:33 PM »
Thanks a lot Prof. Govindjee. Could you please comment on parallelization of the mesh generation algorithm. Between mesh generation and solution of the  linear/nonlinear equations (using PETSC), which one is the bottleneck ?

I have written a FEM code using PETSC where I am now generating a simple 3D brick element mesh. Now I want to write an automated mesh generation program in parallel. I am wondering if an automated mesh generation program in more complicated geometries could be a bottleneck to parallelization of the whole FEM calculation process.

Prof. S. Govindjee

  • Administrator
  • FEAP Guru
  • *****
  • Posts: 1164
Re: Parallel performance of 3D FEM codes
« Reply #3 on: June 20, 2013, 08:58:12 AM »
Equation solving is the dominant bottle neck.

kashraf

  • New Member
  • *
  • Posts: 3
Re: Parallel performance of 3D FEM codes
« Reply #4 on: June 20, 2013, 11:11:17 AM »
Thanks a lot. This is really useful.