Question about SU2 convergence
 

Hello guys,

In fact, I have a big problem with SU2 now. It’s about metis and convergence. When I use mpirun to run SU2 with different -np numbers, the speedup curve is not an increasing function.

For example, when -np = 42 the TestCase runs for thousands Iters with the residual around -4. But when I run SU2 with -np=84 the Testcase exited succeed for 570 Iters, the residual around -1. Why?

Does the SU2’s kernel algorithms have different convergence speed with different metis diced?

Or may be I got the wrong way?

Thank you

Hi,

While the values of the residuals may not be equivalent iteration-by-iteration for implicit, parallel calculations, the residuals should report identical values for every iteration if you choose an explicit method.

Unfortunately, it is difficult to identify some problem/mesh-dependent issues. Just to test your installation without any aggressive settings that may be causing issues, you might turn off multigrid (MGLEVEL= 0), switch to explicit time integration (TIME_DISCRE_FLOW= EULER_EXPLICIT), and turn the CFL number down (perhaps CFL= 0.5). Try running several iterations, say 10, on various numbers of processors. If you are still seeing odd behavior and the residuals don't match with each iteration, I recommend that you check your configure options & rebuild the code.

Hope this helps,
Tom

Thank you very much, that may be help me. By the way, I wanna to compile SU2 to run it faster. Is that some best practices? Or may I use -O3/-O2 to make it better? Because if I use O3 the precision may get down.

Thank you