[chiven]

As we know, OpenFOAM is an amazing CFD software. However, these time, I meet the problem about how to improve the calculation efficiency. I use OpenFoam to solve multi-phase flow. In my case, there are about a million mesh volumes. The time step is 0.00028s to ensure the maximum Courant number of 0.2. Parallel calculation with 64 processors is performed. However, 18 Execution Hours can only finish 1.15s time. About 20s are necessary for my case, this means about 300 Execution Hours will be needed for a case. This is a terrible number.
Has anybody used OpenFOAM to finish the large calculation? Let's discuss how to improve the calculation efficiency.
Best regards,
Chiven

[matejfor]

Hi,

well according to my experience, this does not need to be a problem of OpenFOAM itself. Have you checked the efficiency of the computation? How mush of the time is the actual computing and how much is the node-node communication? If you divide 1e6 by 64 processors you have ~15e3 cells per node. Maybe too many processors. It will be faster, than on fewer nodes only if the paralel speed-up beats the communication overhead.

On the other hand, if you know how to speed-up OpenFOAM, everyone would be gratefull.


matej

[chiven]

Originally Posted by matejfor

Hi,

well according to my experience, this does not need to be a problem of OpenFOAM itself. Have you checked the efficiency of the computation? How mush of the time is the actual computing and how much is the node-node communication? If you divide 1e6 by 64 processors you have ~15e3 cells per node. Maybe too many processors. It will be faster, than on fewer nodes only if the paralel speed-up beats the communication overhead.

On the other hand, if you know how to speed-up OpenFOAM, everyone would be gratefull.


matej

Hi, dear matej, thank you for your comments. I have ever seen a thread and know that ~15e3 cells per node are recommended.
In my case, the execution time and clock time are 9576.23 s and 10778 s, respectively. Maybe it does have a space to up.
I am wondering how to speed-up the calculation.
Best regards,
Chiven

[markusrehm]

Hi Chiven,

it is very difficult to give general numbers and answers for decomposition. It depends on the solver, the discretization, the interconnect, CPU, ....

So I would recommend to run benchmarks and plot the result against #cores (1,2,4,8,16,32,64). Then you'll see where your problem runs more efficient and you can decide.
You could also try different modes of decomposition to minimize the communication overhead and check different linear solvers.

Regards, Markus.

[chiven]

Hi, dear Markus, thank you for your comments. I shall try your sugestions. These days, I am struggling for this point.

Best regards,
Chiven