[TahminehAd]

Hello Everyone,
Is it possible to have different results of a transient simulation on a personal laptop and a cluster with 28 cores?
I have this problem that I run one simulation on my laptop and on the cluster to speed up the simulation time, but the results of the same file, with the same simulation set up are different. For example, on my laptop the courant number is relatively small while on the cluster it is so much high. Or on my laptop the simulation converges with lower number of iterations but on the cluster it doesn't converge or in high number of iteration the convergence happens.
I want to know if that is even possible and if yes under what conditions this problem happens.
Thank you in advance.

[ghorrocks]

This is uncommon, but it does happen. It is usually caused by a strong gradient (eg a free surface, shockwave or other very severe gradient) lying on top of a domain partition boundary.

The fix for this is to change to a different partitioning algorithm which keeps the mesh partition boundaries away from the strong gradients, or if it needs to cross them crosses them around 90 degrees.

So if you have a free surface which is horizontal, if you use vertical partitions it will be much better than horizontal partitions.

The default partitioning algorithm is METIS, so try one of the recursive bisection or directional methods.

[TahminehAd]

Originally Posted by ghorrocks

This is uncommon, but it does happen. It is usually caused by a strong gradient (eg a free surface, shockwave or other very severe gradient) lying on top of a domain partition boundary.

The fix for this is to change to a different partitioning algorithm which keeps the mesh partition boundaries away from the strong gradients, or if it needs to cross them crosses them around 90 degrees.

So if you have a free surface which is horizontal, if you use vertical partitions it will be much better than horizontal partitions.

The default partitioning algorithm is METIS, so try one of the recursive bisection or directional methods.

Thank you for your answer. Exactly after I added free surface into my simulation, that problem happened. So instead of METIS algorithm there are four alternative methods: MeTiS Recursive Bisection, Simple Assignment, Directional Recursive Coordinate Bisection, Optimized Recursive Coordinate Bisection, Recursive Coordinate Bisection. Which one do you think can solve the problem better than others?

[ghorrocks]

You have to choose which ever one keeps the partition boundaries away from the free surface. You can view the partition boundaries by loading a res file into CFD-Post and look for the variable Partition number (or something like that).