[HR1991]

Hi foamers,
I am currently working on simulation of a surface-piercing propeller. I use snappyHexMesh for generating the cells on propeller surface. my problem is the slow simulation speed. I am parallel-runing a case with about 1 million cells using a RANS simulation with k-omega SST model in a computer with following specs:
CPU: intel core i7, 3.4GHz with 6 cores and 12 threads.
RAM: 6 GB DDR3.
my simulation is very slow such that it takes about 48 hours to reach to time t=0.02 seconds, while I need that the simulations reach to time t=0.5 seconds, at least. I am using parallel run with 10 processes.
my question is that, how can I speed up my simulation? is it a good idea to increase the RAM up to 32 GB? what is the efficient amount of RAM needed for grid size of 1 million for aforementioned cores?
Thanks.

[KateEisenhower]

Hi,

a few questions you, I and hopefully someone else can think about:

1) "6 cores and 12 threads" Are we speaking of hyperthreading here? I wouldn't recommend this.
2) The 6 cores are on one socket?
3) How many memory channels do you have? Is the memory equally distributed?
4) As a rule of thumb: One million cells roughly equals 1 GB of memory. So this shouldn't be an issue here.
5) Do you use adjustable time-stepping?
6) Out of curiosity: How did you calculate the t=0.5 s.?
7) Have you tried core-binding?

Best regards,

Kate

ps.: Did you try serial run? What acceleration rate did you achieve with parallel?

[HR1991]

Hi Kate and thank you for your response,
your questions responses are listed as below:

1) yes. I used hyperthreading. but, now, as you mentioned, I think it isn't a good idea. Anyway, I will not use it.
2) yes. the 6 cores are on one socket.
3) the memory is dual-channels and I don't know if it is equally distributed. how can I find it?
5) yes. I used adjustable time stepping.
6) by my previous experiences. I have simulated some cases in coarser grids and I found that when time reaches to t=0.5 s the value of drag and torque coefficients would converge.
7) no, I didn't try it. I will do it and report the result.

I didn't compare the parallel run with serial. I will do it and inform you.

Best Regards,
Hossein

[KateEisenhower]

Hi,

some remarks:

3) I'm sure you can find it out using the terminal, but I don't know how. For sure you could open your machine and find out which RAM-modules sit in which slots.
For example, if you have 3*2 GB of RAM, your simulation might run faster when only using 3 of your 6 cores. Otherwise 2 cores would have to share one memory channel. This could be a bottleneck when using slow DDR3 RAM.

6) I have not much experience with transient simulations. But shouldn't a transient simulation converge for every timestep? Maybe some other user could make this clear?

I'd say your next step should be a simulation on 1 core with besides that unchanged input files. Then we can see if it is a parallel-computing issue at all.

Best regards,

Kate

[HR1991]

Hi again,
thank you for your response,
as I know, a transient simulation convergence in each time step is depended on the number of iterations in pimple algorithm. I mean the nCorrector and nOuterCorrector parameters. although this parameters should have optimized values.
for more clarification, I refer to following link:
https://openfoamwiki.net/index.php/O...hm_in_OpenFOAM

Bests,
Hossein

[KateEisenhower]

Hi Hossein,

this is still a little mystery to me. For clarification I asked Tobi who wrote this very usefule guide:

http://www.cfd-online.com/Forums/blo...m-part-ii.html

See the last comment!

Best regards,

Kate