Does grid decomposition affect simulation speed?

[robson_passos]

Dear, I am preparing a simulation whose domains have the following numbers of grid points:

e_we = 211, 280, 532,
e_sn = 198, 226, 358,

I have two computational nodes available with 64 cpus each, I wonder if the most correct is to use 128 cpus, which decomposes into 8 NTASKS_X and 16 NTASKS_Y or 120 cpus that decomposes into 10 NTASKS_X and 12 NTASKS_Y. The last option more closely approximates a square. On the other hand the first uses more cpus. I wonder if there is a performance difference given this difference in decomposition.

I did some tests and both start the simulations without returning errors. But I didn't have the opportunity to run the two options to compare the speed, as it is necessary to save on resources.

Thank you very much in advance,
Robson

 

[Ming Chen]

Hi Robson,
If you run with all the 128 processors, the domain will be decomposed into 8 NTASKS_X and 16 NTASKS_Y. More processors are supposed to run faster, but note that the communication between processors will take time. There is no definite way to determine how many processors will give you the best scaling. One has to try different numbers and decide which ones is the optimal.