Hello Everyone !!
Gromacs with double precision is argued to be more appropriate for performing REST2 simulation specially for large systems where the total energy happens to be of the order of power 6 or more. Unfortunately, I have such system and even after using 1000 cpu cores and playing around nstlist, PME order etc. I am not getting appropriate speed (Just up to 3ns/day) for standard MD which will be reduced further after inclusion of replicas.
I was considering to use single precision gromacs with gpu can be used in order to increase the simulation speed in case of REST2 and hence needs suggestion and comments over this. Any suggestion, sharing of experience, publication reference and comments etc. will be helpful.
Thank you
This below links might be useful for you:
http://www.gromacs.org/GPU_acceleration
http://www.gromacs.org/Documentation/Acceleration_and_parallelization?highlight=hybridization
You should first of all focus on parallel scaling. The most critical part here is the partition of pme to pp ranks and I found earlier that the default settings of gromcas are very bad it this terms. You control the number of PME ranks with -npme to mdrun. You should select the number of cores in -npme to multiples of the cores per node. From your logs you can see the default settings gromacs chosed and also the pp/pme imbalance. Given the imbalance you need likely to increase -npme.
Also, using a double precision versions seems not a good idea in my opinion and anyway this drops GPU support.
Thank you Normal Geist for you suggestion. Indeed parallel scaling is very very important and I have tried different combinations but still not getting good speed. I think I will have to drop double precision.
Thank you once again.
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running