Attached is the tail of log files from 3 different runs on same system. 1 Node is used to compute using 8, 4 and 1 processor respectively to check the speed
I am running an MD simulation on lammps 30July2016 on 4000 atoms & getting a max rate of 2ns/day (which i consider very slow for this small system).
Here are the details :-
I need help to troubleshoot why simulation rates are slow. With just 4K atoms. Kindly let me know if you need any other information to understand the issue
Yes, 8 processor takes less time compared to 1 processor but for a system this small, my simulation rate is very slow. I do not have access to GPU processing right now.
Any suggestion for this CPU processing to make is speed up?
You're right, 25 timesteps per second (in your best case, 8 processors) is quite slow.
As you can see, more than 80% of the simulation time is spent calculating Kspace, that is, FFT for the long electrostatic interactions.
You can start by optimizing these settings in your configuration file.
How about running the simulation in single precision? Is that acceptable? There would be speed increase there too.
@Fernando, thank for replying but I could'nt follow. I did run simulation on a single processor. Speed decreased with 1 processor as shown in attached image. (or did you mean something else? )
What optimizations could you suggest regarding Kspace ?
Yes, you ran in 1, 4 and 8 processors. 8 processors being the fastest, which is an expected result when there is no communication overhead.
What I meant by optimizing Kspace is to study LAMMPS documentation and check which parameters can you modify to improve electrostatic force calculations speed.
In line with Fernando's suggestion, if you check the lammps documentation on KSPACE package you will see that FFT is necessary for performance. Is your lammps built with FFT package? That would have a direct impact on performance.
http://lammps.sandia.gov/doc/kspace_style.html
Also you should check for the accuracy parameter. If too tight, then the simulation will be time consuming.
Finally, through kspace_modify command you might be able to make further adjustments.
http://lammps.sandia.gov/doc/kspace_modify.html
Cheers!
How to use GPU using CUDA? If anyone have installation steps, then please share.
- Badges
- Science topic