CFDEM calculatiom time 25E+3 particles & 1.4E6 cells??

PaulWinkler's picture
Submitted by PaulWinkler on Sun, 05/19/2013 - 16:48

Hi folks,

in my last project I used LIGGGHTS to simulate a compression test up to 25000 particles with a diameter in a range of 0.02 to 0.0011 m in a cylinder (r=0.07m, h=0.07m). I did the simulation in steps, the longest last 3 seconds, the calculation takes 5 days on a single core (Used to calculate other simulation in parallel on the other cores, what was in total faster). Now I want to extend the model with CFD coupling and want to choose the cell size in the size of the smallest particles. So there are round about 1.4E+6 cells in the grid. (I am a beginner to CFD, maybe I am totaly wrong, please give me an advice.) How long will this take if I use one core for DEM and one for CFD? What is more CPU consuming? Is there an easy way to setup a CFDEM with GPGPU support? After reading a while, all I got was a clear maybe... Seams to be the best way to provide an own Linux/GNU distribution for the hole thing.

Thanks a lot.

Regards

Paul

Forums:

evansmuts | Tue, 07/02/2013 - 15:40

Hi Paul

I see this thread is a little old, but I thought I would share some things I found from my simulations. Hopefully my comments help:

1) CFD cell size >= 3 particle diameters. So you can probably reduce your number of cells...

2) DEM scales very well in parallel. I am currently running 14000 particles, and they run great on 12 cores. Performance might be better on 16, but there are very few 16 core CPU around here.

3) CFD also scales well with processors. Having around 100k cells on 1 processor should be fine (been a while since I tested that). Though more than that and you should go parallel. I use 1800 cells. For such a small number of cells, CFD actually gets worse for increasing number of cores. But the improvement in DEM performance is worth it.

4) CFD-DEM coupling does not scale well when going to parallel. For me, CFD and coupling takes roughly 50% of timestep (14k particles + 1800 CFD cells). My CFD-DEM timesteps take about 0.05 seconds to compute on 12 cores (according to the log file).

5) I haven't run on 1 or 2 cores in long time. But I tried my simulation on 4 core machine other day for a test. It took a VERY long time to compute.

Cheers
Evan