The current CVS of Adun now automatically threads nonbonded and Generalized Born calculations to multiple cores if they are available.
The attached graph shows the speed-up for Gas phase and GB simulations of a 6500 atom protein. Since the threading only applies to certain calculations you don’t get a linearly scaling speed-up. Instead the expected speed increase depends on the percentage time spent in the threaded calculations in serial – so bigger systems will show better scaling.
For this test case threading gives around a 4-5 times speed increase on six cores e.g. from 0.48 to 0.1 secs per step for 14 angstrom cutoff gas-phase (x4.8) and from 2.78 to 0.66 (x4.2) for 14 angstrom GB. The reason the GB calcs show less improvement with 4+ cpus seems to be related to memory access issues and we’re currently looking at how this can be improved.