[vaina74]

On these days I'm running some simulations with MRFSimpleFoam solver, concerning a marine propeller. I tried different method of decomposition and I saw that metis is a bit faster than simple one (along rotation axe). I also tried to test a different use of cluster noded (I can use 1, 2 or 3 8CPU-nodes on a cluster). I'm very surprised to find out that simulations are faster on 1 single node (with 2 nodes a single iteration takes almost a double time). I know that communication among more nodes can be difficult but all this difference looks too much to me). Are there any hints or specific settings in decomposeParDict to have a better running? Maybe does it depend on the globalFaceZones? This is my decomposeParDict:

numberOfSubdomains 8;

method metis;

globalFaceZones ( cyclic_1Zone cyclic_2Zone symmetry_1Zone symmetry_2Zone propZone domainZone );

metisCoeffs {}

[vaina74]

Please, any idea? I checked the decomposition out: the cells are equally shared by CPUS. Can 'large' areas of globalFaceZones affect the parallel computing and make a 1-node simulation faster than a 2- or 3-nodes one? I can't understand all that difference in time consumption.