In e-mail from Andrew Yool which is contained in Colin Jones's e-mail from Fri 5 September.
| Number of PEs | ||
|---|---|---|
| 128 | 256 | |
| Factor increase to NEMO of adding MEDUSA | 2.73 | 2.10 |
Currently CICE is restricting the number of PEs that we can use with NEMO and MEDUSA.
Estimates of run time for NEMO+MEDUSA are wanted for the upcoming meeting, but we don't have a stand-alone NEMO+MEDUSA, i.e. without UM (we have requested one and there's an ongoing conversation with Andrew Yool and Julien about this). What do have is the information from above combined with timings for ORCA025 (on Ocean only page), and Yongming's coupled timings.
The graph of ORCA025 performance on the `Ocean only' page suggests that 128 PEs runs at about 1.4 years/day, so if MEDUSA increasing the run time by a factor of 2.73 then this suggests 1.4 / 2.73 = 0.51 years/day. And for 256 PEs, the graph of ORCA025 performance suggests about 2.6 years, so for MEDUSA 2.6 / 2.1 = 1.23 years/day.
These suggest the three results below. We can fairly sure that in the UKESM0.1 runs with 16 * 23, and 24 * 23 PEs for NEMO+MEDUSA that the NEMO+MEDUSA component is the slowest component - but it's largely a convient guess the 32 * 25 PEs run.
| Source | Total PEs | Time for one model year (using next column) | Speed (model years per day) |
|---|---|---|---|
| Estimates using ORCA025 graph and information from Andrew Yool's e-mail | 128 | 0.51 | |
| 256 | 1.23 | ||
| Yongming's UKESM0.1 timings | 16 * 23 = 368 | 100,465s | 0.86 |
| 24 * 23 = 552 | 75,789s | 1.14 | |
| 32 * 25 = 800 | 62,609s | 1.38 |
Clearly the estimates based on graph and Andrew's e-mail and estimates from Yongming's UKESM0.1 runs don't correspond, because otherwise 256 PEs is faster than 368 PEs. The estimates from Yongming's UKESM0.1 runs are likely to be more acurate.
E-mail from Richard on 23 July 2015 gives
| Total PEs | Speed (model years per day) | Time for one month (s)* |
|---|---|---|
| 32 | 2.9 | 2,483 |
| 64 | 5.7 | 1,263 |
| 99 | 8.7 | 828 |
| 192 | 14.0 | 514 |
E-mail from Julien on 30 July 2015 gives
| (NEMO East-West)*(NEMO North South) (total) | Model run length | Total time (s) | Speed (model years per day) |
|---|---|---|---|
| 16x12 (192) | 10 days | 6,322 | 0.380 |
| 32x12 (384) | 10 days | 2,760 | 0.870 |
| 32x20 (640) | 10 days | 1,578 | 1.52 |
| 32x30 (960) | 10 days | 1,082 | 2.22 |
