Scheduled Downtime
On Tuesday 24 October 2023 @ 5pm MT the forums will be in read only mode in preparation for the downtime. On Wednesday 25 October 2023 @ 5am MT, this website will be down for maintenance and expected to return online later in the morning.
Normal Operations
The forums are back online with normal operations. If you notice any issues or errors related to the forums, please reach out to help@ucar.edu

How many cores make CAM5-SE faster than CAM5-FV

O

oyvind_seland@met_no

New Member
 I have tested CAM5.3 with our local extensions (Slightly higher complexity than CAM5.3-MAM4 ) on a NeXtScale nx360M5 WCTThe extensions slow down the model with around 15 % At 7680 cores the turnover of CAM-FV quarter degree resolution is 2.36 years / day.  The equivalent for CAM5.3 -FV would be 2.7 years/day.However CAM5.3 SE yields less than 2 years for the same set-up. Now as far as I can see 7680 is the maximum number of cores for CAM5.3 FV so SE should be faster for a higher number of cores. I could not test SE with higher number of cores since I had only a very limited amount of CPU hours available for my tests.
Is SE only faster due to a higher possible number of cores or are there any tricks I should know of to make the SE version faster?   
 
S

shian-jiann_lin@noaa_gov

New Member
My previous attemp of posting failed. here is my take.
I am the original developer of FV core used in the CAM-FV, which is based on lat-lon grid. Its scalability is limitted by its use of polar filter, which is the main reason we at GFDL developed the next generation FV core on the cubed-sphere, called FV3 (FV cubed).
Due to FV's inherent advantage in CFL condition (larger time step) and more efficient tracer transport algorithm (with 2 decades of continued optimization!), it is extremely difficult for the SE core to outperform the FV lat-lon when there are over 20 tracers (as in CAM5). To beat the Fv lat-lon, you wil need to use FV3. There is a version of FV3 within CAM. It should outperform the lat-lon FV at pretty much any resolution and at any core core that is higher than 500 cores. At 7K or 8K cores, the FV3 should be 5 to 10 times faster than the old FV lat-lon (and the SE core). 
 
O

oyvind_seland@met_no

New Member

Code: 
It has been a long time since I looked at the tread, but at last thank you for the information. Do you have examples on how to set up namelist / Makefile etc in <br />order to test FV3 for updated version of CAM / CESM?

Code: 
Best regards

Code: 
 

Code: 
Øyvind Seland

Code: 
Norwegian Meteorological Institute
 
You must log in or register to reply here.
Top