Hi everyone! I'm starting to learn how to perform Replica Exchange Molecular Dynamics (REMD). At this point, I'm interested in perform those simulations in GPUs, but the resources I have are limitated for that since the number of replicas i want to use is greater than the number of GPUs I have. Any experience using multiple replicas in a single GPU?
I think this is usually no problem, altough it depends on the MD code. For NAMD as an example, the restrictions are more on havin at least a CPU per replica, but they can share GPUs without issues (except of performance).
What kind of system do you want to sample with REMD?
No specific experience here, but as Norman said, it depends on the code.
For AMBER you may check this thread:
http://archive.ambermd.org/201401/0315.html
In a nutshell: you can, but you need a lot of memory and performance may not be optimal, and your GPU has to be configured in non-exclusive mode.
Note:
For the sake of completeness I must add you can run your REMD simulations independently (like regular MDs) and then program the swapping yourself, and then re-send the REMD simulations like regular MDs manually, and repeat.
To program the swap you can use the metropolis algorithm (in AMBER19 manual you can check section "21.5.4. Running Temperature REMD simulations").
Take into account that this method won't be useful if you have to swap very quickly (because of the extra time it takes to start a MD). But for loose calculations (like swapping every 10ns) it may work.
- Badges
- Science topic
- Similar topics
- Chemistry
- Theoretical Chemistry
- Computational Chemistry
- Molecular Dynamics