[cnqin]

How to set up the parallel computing of CFX to make the calculation speed faster? I found that it does not seem that the more cores are set, the faster the calculation speed is. Is this related to the number of grids?

[ghorrocks]

Setting up CFX to run parallel is described in the CFX documentation.

The CFX parallel algorithm is pretty good and most simulations will speed up with a small number of partitions. But once you have more than a small number of partitions (maybe about 12), you need a computer which is designed carefully for parallel operation or you will not get much of a speed up.

There are also some physics models which do not run very well in parallel, the main one being Monte Carlo radiation modelling.

[Opaque]

Theoretically, parallelization is associated with infinite speed-ups, but in reality, it is not and it is bounded by several factors.

If you have 10 different tasks, you cannot divide them by 1000 regardless of whether you have the resources. Imagine you are assembling a device with multiple components, there will be a limit on the number of people that can simultaneously install parts either because of space, order dependency, or amount of parts.

In CFD, the list includes:
1 - Hardware limitations: More cores in a single computer are bounded by the memory bandwidth available to a single core, and the speed transfer between cores when transfer is needed. If the core associated with your task needs to store data outside its direct access, the best algorithm cannot overcome hardware slowdowns.
2 - Number of elements/vertices per core: at some point, the effort to transfer data between cores becomes so high that the addition of another core negatively affects the performance -> Communication overhead.
3 - Convergence: depending on the solution algorithm chosen, the more partitions added, the more explicit the solution algorithm becomes --> unstable or longer to converge. The additional cores though faster to execute the task, the contribution to the overall convergence is smaller; therefore, more iterations to converge -> slowdown.

Those are 3 from the top of my head. The more you fix them, then others start to show up or become more visible.