[wrik]

Hi,

I am running SU2 Quickstart NACA0012 Mach=0.8 steady case, on a 48 core Linux machine. I have built from source code for parallel using proper ./configure options. OpenMPI is loaded which provides me with mpicc and mpicxx.

I am suing the python command:
parallel_computation.py -f inv_NACA0012.cfg -n 8

to use 8 cores which takes about ~1500 seconds to converge. Using 2 cores takes about ~170 seconds. The results however, match for both the cases.

I am not sure why using more cores takes more time. Is it because that the problem is so simple that domain decomposition is taking more time than solving and thus running in parallel is of no use here? I have attached config.log and config_CFD.cfg file which will show you that MPI support and paraMETIS support is provided, and the domain is indeed partitioned. Kindly provide some suggestions.

Wrik

[hlk]

Originally Posted by wrik

Hi,

I am running SU2 Quickstart NACA0012 Mach=0.8 steady case, on a 48 core Linux machine. I have built from source code for parallel using proper ./configure options. OpenMPI is loaded which provides me with mpicc and mpicxx.

I am suing the python command:
parallel_computation.py -f inv_NACA0012.cfg -n 8

to use 8 cores which takes about ~1500 seconds to converge. Using 2 cores takes about ~170 seconds. The results however, match for both the cases.

I am not sure why using more cores takes more time. Is it because that the problem is so simple that domain decomposition is taking more time than solving and thus running in parallel is of no use here? I have attached config.log and config_CFD.cfg file which will show you that MPI support and paraMETIS support is provided, and the domain is indeed partitioned. Kindly provide some suggestions.

Wrik

We generally recommend not going lower than 5,000 points per processor in parallel problems. This test case is very small, with only a little over 5,000 points total, so I would actually expect it to run the fastest on a single core, or at least not much slower than 2 cores. The reason is not only the domain decomposition, it is also that when running in parallel there is an additional cost associated with communicating information between the different cores. For example, in this case when using 8 cores, at each iteration each core only has ~625 points to compute, but then has to communicate with each of the 7 other processors prior to the next iteration.

[wrik]

Thanks for the suggestions! I have tried parallel run on more complicated models and they just work fine.

Wrik