ideal processor layout

[climatemodelling@gmx_com]

Dear Forum, I am running AGCM-only and AGCM-OML experiments on the 1.9x2.5 finite volume grid. The machine I am using is organized by having 24 processors on each node.So far I tried the following processor outline utilizing only one node: (24/1/0) for ATM,LND,ICE,OCN,CPL and (1/1/0) for GLC. I get about 3.5 model integration years per 24 hours.My question is what the best processor outline for the individual components would be to maximize the speed on say either 2,3 or 4 nodes (24 processors each) using CAM-only or CAM+OML. I would also be happy to use T42 instead of the finite volume core if that would result in a considerable increase in speed? Thanks! 

 

[jedwards]

Hi,

The Machine hopper which is listed in the Machines directory has 24 tasks per node and you might gain some insiight by looking at the configuration you would get on that system. However there are a lot of factors that will affect cesm performance and your best insights are going to be gained by running a few scaling tests on your system to see what works best.

 

[climatemodelling@gmx_com]

Hi and thanks for the reply!
I checked the hopper pes outline in the user guide and it seems that the same number of processors (N/1/0) for (NTASKS/NTHRDS/ROOTPE) for all model components is used in the example.
I was wondering if there is a general guideline or rule-of-thumb for CAM-only and CAM-OML runs of other processors distributions among the individual components (e.g. more processors for the atmosphere component than the data ocean) that the community has found significant speed increases for. As a model user it is hard to imagine where the most likely bottlenecks are.
Thank you very much in advance!

 

[jedwards]

For the AGCM using all of the processors for all of the components is the best setup. So just increase that number to 48,72 and 96. If by AGCM-OML you mean a B compset with active ATM and OCN components then you want to separate the ATM and OCN components so that they can run concurrently. Probably 12 or 24 processors for the OCN and the rest for ATM.

 

[climatemodelling@gmx_com]

Thanks for the clarification!
With CAM+OML I mean the E compset. I assume that less processors are required for the ocean component in the mixed layer mode than for the fully dynamic ocean model (B compset)?