Re: Can large number of atoms define in colvariables reduce the computational speed in adaptive biased force (ABF) simulation via lowering the GPU utility?

From: sunyeping (
Date: Thu Apr 27 2017 - 00:34:49 CDT

Dear Giacomo, 
Thank you for your reponse. Sorry that I didn't make my question clearer. Actually the protein I simulate contains five chains (Chain A, B, C, D and E). Chain A,  B and C form one subunit (subunit I) , and chain D and E form the other subunit ( subunit2). The two subunits bind to each other as a whole. If I just select one atom from the two subunits, for example, I select one atom in chain A and select another atom in chain D, and simulate the distance between them with ABF, I guess chain A may seperate from chain B and C, or chain D may seperate from chain E during the simulation, which can may the problem unneccessarily complicated. So what is the proper way to define the atom groups in the colvar for the distance between the two subunits if including all atoms of them is not neccessary?
The GPU utility is gotten by the command "nvidia-smi", if ABF does not depend on GPUs, then it may not be considered.
Best regards.
Yeping ------------------------------------------------------------------From:Giacomo Fiorin <>Time:2017 Apr 27 (Thu) 02:34To:namd-l <>; 孙业平 <>Subject:Re: namd-l: Can large number of atoms define in colvariables reduce the computational speed in adaptive biased force (ABF) simulation via lowering the GPU utility?
Hi Yeping, if the purpose is to define the centers of mass of two protein domains, you don't really need to include all the atoms.  You can check for yourself that those centers (and the distance between them) don't change much if you select 1 out of 10 atoms, or e.g. the C-alphas.

Also, the Colvars version bundled with 2.12 includes several optimizations, including parallelized calculation of centers of mass.

Lastly, it is very hard to understand what you mean by "GPU utility".  All I can say is that the Colvars code runs entirely on the CPU, so it can benefit from GPU usage only indirectly.  Out of all the features of NAMD, only the most widely used are ported to the GPU.


On Wed, Apr 26, 2017 at 1:51 PM, sunyeping <> wrote:
Dear all,
I am trying to do ABF simulation on a protein with a workstation with two c2050 GPU. The colvariable I choose is distance. If the distance is the distance between two atoms (group1 and group2 both contain one atom), the computational speed seems to be normal (about 0.3 day/ns). However, if the distance I define is that between the centers of mass (COM) of two protein subunits (group1 and group2 both contain about 6000 atoms, the computational speed becomes very low (2 day/ns). When I check the GPU ultility, I find the two atom distance simulation has a relatively higher GPU utility (60%), but the two subunit COM distance simulation has a very low GPU utility (4%).     How can the atom numbers defined in the distance colvariables affact the GPU utility? Is there any method to improve this GPU utility and computational speed in the COM distance ABF simulation?
Yeping Sun
Best regards.

Giacomo FiorinAssociate Professor of Research, Temple University, Philadelphia, PA
Contractor, National Institutes of Health, Bethesda, MD

This archive was generated by hypermail 2.1.6 : Sun Dec 31 2017 - 23:21:14 CST