[Simbelmynë]

Hey Foamers,

I have added a simple scalar transport to interFoam (let's say it is Temperature). It runs well in serial. However, when running in parallel it diverges if the diffusivity is too high (but works well if the diffusivity is low).

Simple or scotch decomposition both diverges although at different points in the simulation. Similarly, if using a different number of processes (other than one), it diverges at different points in time.

Any thoughts?

OpenFOAM v9

[dlahaye]

Please make sure that sure that linear system is solved to same accuracy independent of the number of processors.

The block Jacobi preconditioner that openFoam implements is known to be less performant on larger number of processors (cfr. literature of iterative solution methods, see below). More iterations of the Krylov acceleration are thus required to reach the same level of accuracy.

Good luck.

References

• https://www-users.cse.umn.edu/~saad/books.html
• ddm.org

[Simbelmynë]

Thank you, switching from PBiCG to PBiCGStab or a smooth solver solved the issue.

[dlahaye]

Really? Why is that?

[Simbelmynë]

Originally Posted by dlahaye

Really? Why is that?

I have no idea. All solvers have the same convergence criteria (absolute 1e-16) and they all perform nicely for a few hundred time-steps. Depending on how the domain is decomposed, the PBiCG solver crashes at different points. The other solvers mentioned do not. Some message passing problem perhaps?



All solvers behave nicely in serial.

[dlahaye]

Thanks for the additional information.