- Contemporary messages sorted: [ by date ] [ by thread ] [ by subject ] [ by author ] [ by messages with attachments ]

From: Jason Swails <jason.swails.gmail.com>

Date: Thu, 26 Mar 2015 10:25:55 -0400

On Thu, 2015-03-26 at 11:59 +0000, Glenn Carrington [bs10g3c] wrote:

*> Thanks very much for the information.
*

*> I initially used the Langevin thermostat as per the AMBER tutorial,
*

*> but it took ages to get 100 steps on the GPU. Switched to Berendsen
*

*> and the GPU simulation ran noticeably faster. I have not got the
*

*> timings to hand, but I'm guessing a x2 speed increase at least.
*

I find that surprising... The complexity of generating the random

numbers is O(N) (really 3N, with somewhat-expensive logarithms and

trigonometric functions required to convert a uniform distribution into

a normal distribution), and it is all done on the GPU directly. By

contrast, the nonbonded calculation is O(N^2) (times two, since the GB

radii and GB energies must be computed in separate double-loops). For

large systems, the nonbonded loop should completely swamp the time

requirements of the random number generation. Scott -- am I missing

something subtle here?

That said, vrand=1000 isn't a bad setting... That will require ca. 1000x

fewer random numbers than the Langevin integrator (although I still

suspect this is very small compared to the cost of the GB force

calculation).

*> However, as mentioned I don't think I can use this thermostat with
*

*> implicit solvent. As the structure has 47,000 atoms, do you think the
*

*> difference between Anderson and Langevin will be more pronounced due
*

*> to the system size?
*

Andersen and Berendsen should be very similar in terms of computational

cost with a sufficiently large vrand. That said, 47,000 atoms is a

*huge* system to be running with GB... very ambitious.

*> I'm hoping so, as the simulation will be painfully slow otherwise.
*

*> Also do you know what is a good setting for vrand?
*

*> I'm not familiar with using nmropt to guide the equilibration temp.
*

*> Hope you don't mind me being cheeky, but could you possibly cut and
*

*> paste a sample of your input file to provide some guidance? Don't
*

*> worry about it if it is inconvenient though.
*

I have a repository of input files that I put on Github, complete with a

README file that hopefully describes my naming convention:

https://github.com/swails/Mdins

"slow_heat" is the nmropt-controlled temperature variation.

HTH,

Jason

Date: Thu, 26 Mar 2015 10:25:55 -0400

On Thu, 2015-03-26 at 11:59 +0000, Glenn Carrington [bs10g3c] wrote:

I find that surprising... The complexity of generating the random

numbers is O(N) (really 3N, with somewhat-expensive logarithms and

trigonometric functions required to convert a uniform distribution into

a normal distribution), and it is all done on the GPU directly. By

contrast, the nonbonded calculation is O(N^2) (times two, since the GB

radii and GB energies must be computed in separate double-loops). For

large systems, the nonbonded loop should completely swamp the time

requirements of the random number generation. Scott -- am I missing

something subtle here?

That said, vrand=1000 isn't a bad setting... That will require ca. 1000x

fewer random numbers than the Langevin integrator (although I still

suspect this is very small compared to the cost of the GB force

calculation).

Andersen and Berendsen should be very similar in terms of computational

cost with a sufficiently large vrand. That said, 47,000 atoms is a

*huge* system to be running with GB... very ambitious.

I have a repository of input files that I put on Github, complete with a

README file that hopefully describes my naming convention:

https://github.com/swails/Mdins

"slow_heat" is the nmropt-controlled temperature variation.

HTH,

Jason

-- Jason M. Swails BioMaPS, Rutgers University Postdoctoral Researcher _______________________________________________ AMBER mailing list AMBER.ambermd.org http://lists.ambermd.org/mailman/listinfo/amberReceived on Thu Mar 26 2015 - 07:30:02 PDT

Custom Search