Parallel and Serial Simulations
 

Dear friends,

I was simulating fluid flow inside a duct with a parallel solver and after reaching the convergence condition, I continued the simulation with a serial solver. That was extremely shocking that the result of my simulation changed to a great extent at the very first iterations and remained at the same value. (I was monitoring heat transfer coefficient)

What is the source of this huge difference? Which result is more trustworthy? Incidentally, I changed the mesh a couple of times and unfortunately there was the same problem again.

Any idea is appreciated.

BR
Sasan Ghomi

It is rare that there is a significant difference between the parallel and serial solver, but it does happen. It particular happens in simulations which have very large gradients in a small region of the domain, such as free surfaces and shock waves. When a free surface or a shock wave lies on a partition boundary you can get convergence problems and it fail to converge.

The cause of this is that each partition is solving the volume assigned to it, and there are partition boundaries at the edge to couple the different partitions together. But this means the coupling of the variables between partitions is weaker than the coupling inside a partition - and this means that if you have a simulation with difficult convergence this might be the thing which makes it fail to converge.

So which one is the right answer? The serial simulation has no partition boundaries and so does not possess this weakness, so it is more likely to be correct. But if you are getting differences between serial and parallel it suggests your simulation has poor numerical stability and you better check that your convergence is adequate anyway. It might be a sign of a more important problem (that is, not tight enough convergence).

If properly converged, both simulations must produce IDENTICAL results.

There should NEVER be any difference in a useful solution between serial and parallel runs.

As Glenn suggested, serial runs are more likely to be correct since it is not sensitive to any parallel coding or algorithm issues.