Comparing predictions with coupled jobs

Predictions are being done by selecting the submodels and specifying the speed, e.g. for the first example 

python modelPred.py --submodels=atmosOnlyN96,nemoCiceOrca025 --speed=1.86 --plot=0

Yongmin's run times

At time of writing, we don't have times for MEDUSA and I haven't finished all the ocean runs (so ocean data is likely to change slightly in the future). However with all the data from Tim Graham, we should be able to test whether my python code can correctly predict the speed of Yongmin's runs.

Model Speed (model years per day) Speed/0.9+ (model years per day) Prediction Prediction for speed/0.9 Actual
ATMOS coresOCEAN cores ATMOS coresOCEAN cores ATMOS coresOCEAN cores
atmosOnlyN96 + nemoCiceOrca025 1.86 2.07 60 180 66 200 8*9=72 16*23=368
offOxN96 + nemoCiceOrca025 1.25 1.39 80 120 90 135 8*16=128 16*23=368
offOxN96 + nemoCiceOrca025 1.81 2.01 120 170 136 192 12*16=192 16*23=368
fullChemN96 + nemoCiceOrca025 0.92 1.02 176 80 198 96 16*12=192 16*23=368
fullChemN96 + nemoCiceOrca025 1.53 1.70 324 144 374 160 16*23=368 16*23=368
fullChemN96 + nemoCiceOrca025 1.77 1.97 392 170 460 192 20*23=460 16*23=368
+if we assume that coupling roughly slows the submodels by 10%.

Summary of comparisons

  • It looks like the cores for the Atmosphere is always the limiting factor, i.e. I don't think the model is ever significantly slowed by the ocean.
  • The predicted number of cores for offOxN96 are much smaller than Yongming used, but we think that Yongming didn't have the improvements we made to offline oxidants (i.e. removal of advecting tracers twice and only reading a particular input file every 5 days).
  • The predicted number of cores are too small. However, if we assume that coupling introduces roughly a 10% overhead (setup and message passing), the predicted cores is fairly good for atmosOnlyN96 and fullChemN96 (matching the numbers in bold). (10% does seem high though!)

Till's run times

Model Speed (model years per day) Speed/0.9+ (model years per day) Prediction Prediction for speed/0.9 Actual
ATMOS coresOCEAN cores ATMOS coresOCEAN cores ATMOS coresOCEAN cores
atmosOnlyN96 + nemoCiceOrca1 (xjpdu) 4.3 4.78 160 12 184 15 198 24
atmosOnlyN96 + nemoCiceOrca1 (xjpdy) 5.7 6.33 252 18 304 20 396 24
atmosOnlyN96 (xkpzo) 4.4 n/a 162 n/a n/a 198 n/a
atmosOnlyN96-dt30 (xkpzp) 5.6 n/a ? n/a n/a 198 n/a
+if we assume that coupling roughly slows the submodels by 10%.

Summary of comparisons

  • Predictions suggest that atmosphere is the limiting factor for both coupled runs
  • All Till's run seem slow in comparison to mine (whether coupled or not)