Load Balancing on CESM1.2(WACCM) - Compset FGEOS_C4WSF_L40CN

[pcampbe2@uwyo_edu]

Hello All,Just a small inquiry.  Does anyone have suggestions for an efficient PE layout when running CESM1.2(WACCM), with the compset FGEOS_C4WSF_L40CN, which is CAM WACCM with sulfur chemistry and the sulfate CARMA model?  The default PE layout is:






 
Does anyone have any experience with the default PE layout as well? Thank you, Patrick Campbell

 

[jedwards]

Hi Patrick,  I would say that there isn't a lot of experience out there with that particular compset.   For an F case it usually makes sense to keep all of the components using all of thetasks sequentially so the only difficulty is determining how many tasks to use and that depends on the dycore and resolution you are using.  

 

[santos]

Since this is an F compset, nearly all of the cost will be in the atmosphere component; there's not much to "balance" in this case. This is an appropriate default for yellowstone; on that machine, the only thing you can really do to increase throughput for this compset is to increase the number of tasks by a multiple of 15 (i.e. to increase NTASKS for every component), since we use 15 pes per node on yellowstone. I'm not sure how much of a benefit you get this way, because I'm not as familiar with the performance of specified dynamics cases.The situation is pretty much the same on any other system, except for a) the best value of the "NTHRDS" variable, and b) making sure that NTASKS is a multiple of the number of processors per node (e.g. a multiple of 12 if there are 12 pes/node).