Performance of hybrid
Reasons why hybrid will get faster
- We've yet to properly load balance the hybrid model
- We think that Jnr needs almost the same nodes as Snr
- We think there's a memory leak in code which strips out UKCA from Snr - it means we can only currently run hybrid for 10 day cycles
- Hybrid code still contains a lot of debugging write statements
- We can't currently use IO server with hybrid
Current performance
| On 56 nodes | Close to top speed | |
| GA + StratTrop N216 | 1.01 model years/day | 1.70 model years/day (160 nodes) |
| Hybrid N216<->N96 | 1.28 model years/day | 2.55 model years/day (224 nodes) |
| Hybrid speed-up | 27% faster | 50% faster |