The number of cores in the I J K direction is 2 3 2, and I have tested 2 4 2 to be a little slower than him. My computer has 24 cores and 32 threads, and as long as I have more than 16 cores in mfix, it will be very slow
You are most likely over-decomposing for such a coarse mesh. There is some overhead for the communication in DMP and if there is not sufficient number of cells, the overhead costs more than the benefit of parallel decomposition. If you are running DEM with a bed of particles near the bottom, you will also experience load imbalance and loss of parallel efficiency. It is possible that 12 cores is the optimum decomposition for your current simulation.
Thank you jeff, if there are 12 cores, how do I allocate the number of cores in the three directions of the IJK?
I would compare a 2x3x2 and a 1x12x1 decomposition (NODESI x NODESJ x NODESK). If you are running a DEM simulation where most particles are in the bottom half, then avoid decomposing in the y direction. You can also manually specify the number of cells in each direction using a gridmap.dat file:
1 Like