Page 1 of 1

Test systems never converge

PostPosted: Sun Aug 11, 2013 1:28 pm
by psrgpu
I'm currently testing CULA library, but unfortunately I am having problem with the most part of my test systems. Not only the big ones, but with the small systems too.

I'm using the software R to compare the results between CULA GMRES(for example) and R GMRES implementation (package pracma).

The R version of GMRES has no problem to find the solution (known solution) for 1e-6 residual tolerance, but the CULA version never converges.

I'm attaching the matrix A (coo format) and vector b and I would appreciate if you could take a look at this system.

Thanks.

Re: Test systems never converge

PostPosted: Mon Aug 12, 2013 5:40 am
by john
Could you be more specific? At minimum we'd need to know your typical restart interval, and typical preconditioners, etc.

Re: Test systems never converge

PostPosted: Mon Aug 12, 2013 7:17 am
by psrgpu
Hi, John.

I have been trying a number of combinations. Solvers, preconditioners, etc.
For example:
Max Iterations: 500
tolerance: 1e-5
Restart: 10
No precond.

In this case, I've got
Residual: aprox. 2.29e-1

Just a little correction. I attached an unsorted matrix (matrix_coo.txt) but before sending it to the solver my program sorts it. I am attaching the correct file now.

Re: Test systems never converge

PostPosted: Tue Aug 27, 2013 4:47 am
by psrgpu
No light at the end of the tunnel for me?