Learning Forum

[FaizanYounas]

Hi there. I want to know how to select the optimum no of cores for a simulation. Like, I have 64 cores right now and I have to run a mesh of 80,000 elements with time step size 1e-6. My simulation runs a lot faster if I give 32 processors to fluent instead of 64. Is there any way to figure out how many cores you should give fluent for optimum performance?

[aitor.amatriain]

The number depends on many factors, but based on my experience it is between 10 000 and 100 000 cells per thread. In your particular case there are 80 000/60 = 1250 cells, which is a very low number.

[Amine Ben Hadj Ali]

Fluent can operate at high efficiency even at very low cell count number. However this will also depend on the models you have and your hardware. I generally think that 100K per core is good average.
I can say 1 Million cells can be ran on 8 to 32 cores. Memory is also important so do not ignore that part!

[ai0013]

To add:
How are these 64 cores distributed? Sometimes depending on the architecture, e.g. having 2 sockets with 32 cores each can perform poorer compared to running with 1 socket. This is because the sockets need time to communicate with each other. Is that your case? Also depending on the grid partition some cores might finish their calculation before others, so they need to wait for them.

There other factors to consider in performance tests, I believe finding the optimum # cores is case-dependent.

[FaizanYounas]

Thank you very much. My 64 cores are in 1 socket. I'm running now 160,000 mesh size on 40 threads and it's working okay. I'll try by reducing the cores and see it's effect on the simulation speed

[Amine Ben Hadj Ali]

­ƒæì