v4.2.0 diverges with 32 processors but converges with 16. v4.0.0 worked with 32
 

I have used SU2 version 4.0.0 a lot in the past, and recently I installed version 4.2.0. When testing out the newer one, I ran into some issues with # of processors causing divergence.

I used a mesh and solver settings I had recently used for a v4.0.0 simulation (which worked fine on a cluster when using 32 processors). Everything seemed fine when running it on my desktop with 6 processors, but when I ran it on the cluster with 32 processors it diverged. Then I tried running it twice more using 16 and 6 processors. Both worked, although the convergence history values were slightly different. I did try building v4.2.0 using a different MPI implementation too (Intel rather than mpich), which also diverged when using 32 processors but had slightly different convergence history values.

Is anyone familiar with this issue? In the past I have solved similar issues by adjusting MG settings, but it seems odd that different SU2 versions with the same solver settings and mesh would behave so differently. I also don't understand why different mpi implementations (Intel, mpich) give slightly different convergence history values.


Thanks in advance for any input.