Parallelization for Large Simulation domain

new developments, future functionalities, hints for further applications of MICRESS
Post Reply
SPB
Posts: 2
Joined: Fri Oct 19, 2018 4:03 pm
anti_bot: 333

Parallelization for Large Simulation domain

Post by SPB » Mon Jan 28, 2019 4:57 pm

Hi Bernd/Ralph,

I am currently trying to check the possibility of simulating the solidification process during casting in order to look at chemical segregation by considering a large domain in the order of mm.

Would parallelisation help in this cause?
As mentioned in one of the forum threads, diffusion solver and stress solver scales up well and speed up can be expected in problems dealing with these solvers. Does parallelization help when we consider larger domain sizes for segregation problem (which would be a diffusion controlled problem)?

I tried performance checks with the Alpha-Gamma example in Micress (T001_Delta_Gamma_dri) using 3 6(default number used in the example), 12, 24 threads. There seems to be no difference with the wall clock and the CPU times. How do I know if all threads are active/used in this case?

Please let me know what is going wrong with this and your comments on the above problem.
Thank you.

Best Regards,
Sai

ralph
Posts: 167
Joined: Wed Apr 27, 2011 4:42 pm
anti_bot: 333

Re: Parallelization for Large Simulation domain

Post by ralph » Mon Jan 28, 2019 6:45 pm

Hi Sai,

there are separate executables for serial or parallel execution.
Of course, you can use both for a serial execution (i.e. one thread), but you will only benefit from multiple threads with the parallelized exe, e.g. MICRESS_par_TQ.
The serial exe is faster than a 1 threaded parallel exe because no parallelization overhead is included. If you do not expect to use parallelized solvers, this is the best choice.

Did you use a parallel exe when running Delta_Gamma? It should display in the header that OpenMP is enabled.

Ralph

SPB
Posts: 2
Joined: Fri Oct 19, 2018 4:03 pm
anti_bot: 333

Re: Parallelization for Large Simulation domain

Post by SPB » Tue Jan 29, 2019 4:10 pm

Hi Ralph,

Thank you for the quick response. I used a serial exe for running Delta-Gamma (OpenMP was shown as disabled in the header).
I did change the default exe to parallel one and noticed the optimal parallelisation when 12 threads are used for Delta-Gamma example.

Does the optimal number (of threads) vary depending on the typer of solvers used in the problem / domain size of the problem ?

Regards,
Sai

ralph
Posts: 167
Joined: Wed Apr 27, 2011 4:42 pm
anti_bot: 333

Re: Parallelization for Large Simulation domain

Post by ralph » Tue Jan 29, 2019 4:28 pm

Hi Sai,

Yes.
Solvers, e.g. diffusion, flow, stress, and problem size influence the speedup you can reach.
At some point your memory bandwith and cache size might also limit the number of threads reasonable to use.

Ralph

Post Reply