Open cecilehannay opened 9 months ago
@briandobbins this new intel compiler run (red) is indistinguishable from the any of the 64 series runs. The 64 series runs are just pertlim experiments from the original run 64. The grey thick line is a ~2 year old case base, which does have a different climate, for reference. Though I think the ocean may have been initialized differently in 26g.
@cecilehannay It can't hurt to run ADF on the first twenty years of 64intel and 64, so we can compare the fields in a lat-lon plot.
@adamrher: I had already done this comparison here: https://webext.cgd.ucar.edu/BLT1850/b.e23_alpha16g.BLT1850.ne30_t232.064intel/atm/ For me, they are the same climate.
@briandobbins Do you think we should run this one longer? I am satisfied that there are no significant differences from ifort to ifx. The only problem is with performance, the ifx run is just over 1ypd slower than ifort.
@adamrher @cecilehannay Great, thank you both so much. This was the hoped-for result.
@jedwards4b I don't think so right now. I think down the line, we may do a short run (5 year?) ensemble to ensure there isn't a statistically significant difference in variance or mean, but that doesn't seem urgent given the above. It's only a consideration given that ifort is EOL soon, and thus it could necessitate a change in compilers during experiments in the future. Happy to hear if you think that's unnecessary, but given the inexpensive cost of a short run it seems worth it to me for the peace of mind.
Sorry @cecilehannay I misunderstood. I didn't realize you already did the ADF comparison. The climate is identical.
@jedwards4b there's 30+ timing files each with a SYPD estimate. Is that enough?
no problemo @briandobbins
@adamrher yes we can stop this one - we will continue to watch development of the new ifx compiler and hopefully they will be able to close this performance gap before they end ifort support.
Purpose:
Comparing standard compiler (
intel
) and the 'new' intel one (intel-oneapi
)Additionally, this run can also be viewed as a perturbation of 64.
Description:
Same as run 64 except:
add
--compiler intel-oneapi
to thecreate_newcase
Case directory:
Locally (if still available):
/glade/p/cesmdata/cseg/runs/cesm2_0/b.e23_alpha16g.BLT1850.ne30_t232.064intel
On github: https://github.com/NCAR/amwg_dev/tree/b.e23_alpha16g.BLT1850.ne30_t232.064intel
Sandbox:
Locally (if still available):
/glade/work/hannay/cesm_tags/cesm2_3_alpha16g
On github: https://github.com/ESCOMP/cesm/tree/cesm2_3_alpha16g
Diagnostics:
Diags (if available): https://webext.cgd.ucar.edu/BLT1850/b.e23_alpha16g.BLT1850.ne30_t232.064intel/
Output:
Raw output on derecho (if still available):
/glade/derecho/scratch/hannay/archive/b.e23_alpha16g.BLT1850.ne30_t232.064intel/
CAM Climos:
/glade/campaign/cgd/amp/amwg/climo/b.e23_alpha16g.BLT1850.ne30_t232.064intel
Contacts: @cecilehannay @jedwards4b, @briandobbins
Extra details: