Hi David,

Thank you very much for reply! 

I have checked the lapack documentation and the respective petsc C-files 
beforehand but it did not help very much. 

I am solving the eigenvalue problem to compute the constant in the trace 
theorem and hence the matrix A is associated with the integral on the boundary. 
I do not only get an error when I increase the number of grid elements but also 
when I decrease the part of the boundary over which I integrate. As I decrease 
in both cases the ratio between non-zero and zero elements in the matrix, I was 
wondering if Lapack might have problems with the fact that A has only very few 
non-zero elements? 

I am actually using KRYLOVSCHUR now and that works well, but I was just curious 
why Lapack causes an error. 

Thanks,
Kathrin






Hi Kathrin,

I'm not sure why lapack would fail here, you could try to look up that 
error "Error in Lapack xGGES 260" in the lapack documentation?

In my experience, lapack works robustly unless the problem is "too 
large": lapack treats the matrices as dense, so you "hit a wall" at some 
point and have to go to krylov subspace eigensolvers in slepc instead. 
But it doesn't sound like your meshes are too large here. It sounds like 
you're having issues for anything bigger than a 5x5 mesh, right?

You could try to use one of slepc's krylov subspace eigensolvers (e.g. 
LANCZOS, ARNOLDI or KRYLOVSCHUR) to see if that works better?

Also, you may already know this, but just in case: You can write libMesh 
matrices out using matrix->print_matlab(), and then read them into 
matlab --- you might find this helpful for comparing the matlab vs. 
slepc results for the larger meshes.

David


------------------------------------------------------------------------------
Managing the Performance of Cloud-Based Applications
Take advantage of what the Cloud has to offer - Avoid Common Pitfalls.
Read the Whitepaper.
http://pubads.g.doubleclick.net/gampad/clk?id=121054471&iu=/4140/ostg.clktrk
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to