Parallelization and tiling

[lcbernardo]

Hi all,

I am a PhD student with an interest in physical oceanography and ecological modeling. Though I am able to perform modeling runs using ROMS, there are still many aspects that I do not fully understand and would like to know more about.

Like most members, I'm trying to find the best configuration for the ROMS model I am running which maximizes computation time and resources. The use of parallel modes and tiling provides some additional flexibility, but I would like to ask if there is a possible performance degradation by using too many nodes or improper tiling. Would there be a recommended limit to the number of nodes to use for a given configuration? And what would be the best recommended practice to determine the I and J tiling scheme? I would appreciate any help.

Lawrence

[kate]

Yes, you can ask for too many processors for a given problem. How many that is will depend on many things - the best way to know for sure is to experiment with your exact configuration for some fixed number of timesteps. Make sure the amount of I/O is representative too.

[lcbernardo]

Hi Kate,

Thank you for your reply. I have been conducting tests as you have suggested, and indeed there comes a point where even doubling the amount of processors hardly has any improvement on performance. I have been going about this blindly previously, so thanks for pointing me in the right direction!

Lawrence

[chysun]

Hi,
I am recently reading ROMS code and try to export some of its intermediate variable,but most fields are tiled because ROMS has a parallel computing method.
I can see there is a routine that collect tiled array in Utility/distribute.F but I still cannot handle it.is there anyone who has experience about this issue ? It is appreciated to hear from you.

Best regards,
Haiyun

[kate]

I would instead copy the examples in say wrt_his.F, which calls nf_fwrite2d.F and so forth. nf_fwrite2d.F does the gather if you want to see how it's done, but why not just call nf_fwrite2d yourself?

[chysun]

Hi kate,
Thanks for your reminder.I will read it carefully.

[flcastej]

Hi,

Six years ago I worked on the same problem as you but for the model COHERENS, trying to find the best number of core to be used. I presented the result "Auto-optimization on parallel hydrodynamic codes: an example of COHERENS with OpenMP for multicore" in a conference and I want to share it with you, just in case it could be useful.

Abstract : http://dis.um.es/~domingo/10/WATERBarce ... stract.pdf

Presentation: http://dis.um.es/~domingo/10/WATERBarce ... tation.pdf

Regards,

-Francisco