CNRM-CM5 simulations on 2 nodes are very expensive. A configuration on one node is much better balanced and less expensive. We present here a test simulation to check that the simulated climate does not depend on the configuration.
We have run an experiment similar to PICTL over 2250-2300 but using only one node on tori. The new experiment called PICTL1N starts from 2250 restart files from PICTL.
Technical caracteristics
- Number of procs per component :
| exp | ARPEGE/SURFEX | NEMO/GELATO | IOSERVER | OASIS | TRIP |
| PICTL | 9 | 6 | 1 | 1 | 1 |
| PICTL1N | 5 | 2 | 1 | 1 | 1 |
- CPU cost for one year (calculated as the average over all years simulated)
| exp | Elapsed Time (s) | CPU Time (s) | Memory (Gb) |
| PICTL | 12600 | 133000 | 49 |
| PICTL1N | 15800 | 98000 | 45 |
Thus a cost reduced by 25% for a 25% longer simulation in real time.
Results
Atlas of PICTL1N (compared to PICTL)
- Main remarks from the atlas
From the atlas, it appears that the mean climate is quite similar (unless for temperature -see hereafter). The diagnostics focussing on climate variability are less similar. As explained in the following on the example of surface temperature, it seems that 50 years is not enough to sample climate variability thus explaining that results are different. This page aims at documenting this !
- Focus on the surface temperature
From the Atlas, we noticed a change in surface temperature in the Nordic Seas and around the Antarctic in PICTL1N compared to PICTL.
The following plot shows the temperature anomaly obtained between the two experiment over 50 years.
The changes in temperature are quite important but located on regions of high inter-annual variability.
The following plot is the surface temperature averaged over the box [68N-78N,10W-10E] (box on the former figure) where the anomaly is the strongest in the Nordic Seas. It can be seen that there is a large multi-decadal variability over this region. The difference between the two simulations does not appear as really significant given this multi-decadal variability. To obtain a smallest anomaly, a longer simulation would have been necessary (but this is quite an expensive test !).
In the Atlas, many diagnostics are different from one experiment to the other. The differences seem to be mainly attributable to the different multi-decadal variability obtained. A "perfect" model should reproduce the results whatever is the number of procs used. As this is not the case for NEMO, there is no way to be sure that the model is not affected by the change of procs.