Closed minghangli-uni closed 3 weeks ago
How many cores are used for wav, 24?
Good question. I just took a look at how a similar configuration is configured by CIME (on the cheyenne
supercomputer):
Without WW3:
With WW3:
I think maybe just give it 24 for now and we can test this properly later.
Thanks @dougiesquire. I've made the change.
Surely we want at least as many cores as the ocean ? i.e. 240
Also, its possible there will be WW config files that need changing too.
Surely we want at least as many cores as the ocean ? i.e. 240
24 cores can be like a starting point. Can we update this later?
Also, its possible there will be WW config files that need changing too.
I am not very familiar with WW configuration. What config files might need changing?
Surely we want at least as many cores as the ocean ? i.e. 240
Why do you say that? I had thought that the scaling was limited by the number of spectral bands (without using an unstructured mesh). How many bands are there?
Also, its possible there will be WW config files that need changing too.
@ezhilsabareesh8?
Also, its possible there will be WW config files that need changing too.
There's no need of any changes in the WW3 config files for PE layout modifications. However it is worth checking the wav.log
during the test run for any changes, I can do that.
Thanks @ezhilsabareesh8
I believe there are 15 spectral bands. From ad-hoc discussions I thought we were expecting it to be a similar resource need to the ocean model. Which is what the research from CIME shows ? ( i.e. they use 300 cores )
Okay, I did some basic tests that show that 24 is definitely not a good choice - this runs slower than prior to these changes despite using 5x the number cores 😅.
I think this requires proper testing. Unless anyone wants to take that on now, I suggest we close this PR (and https://github.com/COSIMA/MOM6-CICE6-WW3/pull/46) and open new ones when we have scaling results to inform our decision.
this runs slower than prior to these changes despite using 5x the number cores
Thanks @dougiesquire. This is not a good choice though. I agree to close this #47 and #46 for now until we have proper scaling results. I will do a scaling test and then open an issue in this repo.
Update PE layout and related cice blocks