3D Modal Analysis (Linear Elasticity)

[dapu]

Hi,

Has anyone tried the 3D Eigen Analysis with the iterative solver? I have a simple cylinder model, it works fine, if I use the direct solver Umfpack or MUMPS. But with the iterative solver I couldn't manage to make the solver converge. I tried different iterative solvers and different preconditioning. Attached is a simple test case. My purpose is to simulate a generator stator with more elements. I tried it with or without mpirun. But same, I cannot make the solver converge. Does anyone has an idea? Many Thanks!

Best Regards
Dapu
From Hannover Germany

[kevinarden]

The iterative solvers do not appear to do well with rigid body motion, free body modes. I was only able to get convergence by using restraints. I tried to restrain the body in a fashion that would eliminate rigid body modes, but not hinder the expected mode shape.

itterative.png

(174.43 KiB) Not downloaded yet

case_res.sif

(2.08 KiB) Downloaded 126 times

[dapu]

Thanks for the effort. That was also my thought. I also do the experimental modal analysis on the test bench. In the real-life, we rarely get a completely free boundary. I will give a try to set some constraints on my generator model.

Best Regards
Dapu

[kevinarden]

Using a eigenvalue shift closer to the first expected non rigid body eigenvalue allowed for convergence but it took over 24 hours.
If direct is available why would you want to use iterative?

itt2.png

(185.87 KiB) Not downloaded yet

case_it.sif

(2.75 KiB) Downloaded 99 times

[dapu]

Hi,

because this cylinder model is only a simple test case. I have a large generator model with more elements to solve. I have already tested it with the mumps solver and with the mesh partition. But it is almost impossible to solve. So I tried to solve it with the iterative solver.

Best regards
Dapu

[kevinarden]

Yes, the direct methods will run out of RAM. However, the iterative methods will take a very long time. You need a high RAM computer.

[raback]

Hi

For elasticity you could try "BILUn", n=1,2,.. as preconditioner. It uses only those entries from incomplete LU that exist in block diagonal of A^n.

If you have Hypre try this, for example
https://github.com/ElmerCSC/elmerfem/bl ... linsys.sif

It uses BoomerAMG to precondition the Krylow methods.

-Peter

[dapu]

Hi Peter,

thanks. I have successfully compiled the elmer with hypre. I will give it try!

Best Regards
Dapu

[kevinarden]

I ran the Hypre and BoomerAMG on six processors, it still took 15 hours, but it did solve the rigid body case successfully. 15 hours is better than the previous iterative solver at over 24 hours. Each process consumed 1 gig each.

[kevinarden]

Compared to 5 minutes using Mumps, each processor took 2 gig each for a total of 12 gig. Twice the RAM requirement but 5 minutes compared to hours. Would be interesting to create a very large swap file to give the system the appearance of large RAM. Impacts performance, but it still may be faster than iterrative.