[gj11]

Dear All,

I have been having some problems with running a simulation on a moving grid. I want to see the effect of a moving aerofoil and I therefore run a simulation without motion first using pimpleFoam and then I stop and restart the simulation with pimpleDyMFoam.
Both simulations are exactly the same in terms of meshing and numerical schemes (obviously with the former not including the dynamicMeshDict or pointDisplacement etc.) and all the velocity (DILUPBiCG) and pressure (GAMG) solvers converge in a similar number of iterations at each time step in both cases. Despite these similarities however, pimpleDyMFoam is taking 82 seconds per time step while pimpleFoam is taking 1 second. Furthermore, the cell displacement solver (GAMG) takes only 1-2 iterations to converge so I don't understand where the huge difference in time is coming from?

I was wondering if anybody else had experienced something similar or has any insight into the solvers which could account for this poor performance?

Many Thanks,
Gareth

p.s. The simulation is run in parallel on 300 CPUs with roughly 30,000 cells per CPUs and the difference between execution time and clock time is not significant - I think suggesting inter-processor communications are not the problem?

[gj11]

Furthermore, results from the static grid simulation compare well with experiments I have conducted so I am confident the simulation has been set up well in general but I must have missed a small numerical detail somewhere when adding the mesh motion....

[blais.bruno]

Some questions which might make it cleared to narrow down the problem.

1 - Do you have frequent IO? Those are dramatic for AMI
2 - Have you checked for a lower number of processor (let's say 0.2 or 0.1 of what you are using right now) if the scaling issue is the same?
3 - Is your AMI interface spread out on many processors? Did you try alternative decomposition strategies?

From my experience, I usually experience something like a 20% slowdown or so, but that was with 12 cores on a single node of a cluster. But from what I have seen, AMI style moving mesh exhibits very poor parallel scaling, but someone might have had different experience from me.

Cheers,
Bruno