[AshwaniAssam]

Dear all,

We have an implicit code for which we have used LUSGS method, based on the use of LU-decomposition algorithm from the Numerical recipe book. When we are trying to use an iterative method like GMRES, for some cases the code is blowing up like hypersonic flow over a cylinder. While for some like compressible flow over bumpit gives exact same residual drop as LUSGS.

We have verified the GMRES implementation against the solution to Linear system of equations. We have used Numerical recipe book again for GMRES algorithm.

Any thoughts what possible is wrong in our implementation or understanding?

Thank you!

[Jonas Holdeman]

I have had trouble using GMRES with incomplete LU preconditioner (actually with the LU preconditioner) in MATLAB. I am quite happy now using the stabilized biconjugate gradient with Jacobi preconditioner. I am not familiar with the LUSGS method.

[sbaffini]

There are several linear algebra packages with the GMRES (petsc, hypre, etc.), I would refer to those for the implementation.

[KingJGod]

Generally speaking, this is probably because of your preconditioner and the accuracy of your numerical jacobian evaluation

[enm]

I think the main problem is a preconditioner or the matrix itself (Jacobian) as KingJGod stated.
Here is how i use it:
https://www.sciencedirect.com/scienc...21999119308940


And there's no problem with gmres or bcgstab.

[AshwaniAssam]

Thank you KingJGod and enm for the answer. I also suspect that. Recently, i have started using LIS library. I will try to use the preconditioner which I was not using earlier. I will report back how things work.


Again thank you, and enm thank you sharing the JCP work it will be very helpful.