Decomposed sHM runs slower than on single thread
 

Hello,
I'm meshing a more-or-less cylindrical geometry (screenshot attached below, flow through a duct) with sHM. I used a simple (2 2 2) decomposition in decomposeParDict ((2 2 1) for 4 threads). I'm running the case in this order:


blockMesh
surfaceFeatures

decomposePar
mpirun -np 8 snappyHexMesh -overwrite -parallel


However, the more cores I use, the slower snappyHexMesh runs. Just adding new boundary patches takes too long to let it finish the job.



I'm using Ubuntu 16.04 LTS on Windows 10. The processor is a 4-core (8 threads) i7-7700HQ. When using 8 threads, the CPU usage jumps to 100%, while with 4 it stays at about 60%, so the decomposition itself seems to be working.


What may be the cause of such behaviour? I'll gladly provide more details as needed.

how many cells per processor?

This is perfectly normal because your processor has 4 real cores so your max parallelization limit is 4!

~3000 after blockMesh, limit in sHM set to 0.8*10^6.

I weren't sure about that, so I tried going over 4 - there are 8 separate CPUs listed in the resource monitor and the total CPU usage approaches 100% only when the case is decomposed for 8 processors.
However, any decomposition seems to be much slower than using just a single core. I've tried using every number of processors between 2 and 8, yet the performance always dropped.