Re: [AMBER] problem using pmemd.cuda with barostat=2 option

From: Jason Swails <>
Date: Wed, 05 Nov 2014 13:47:20 -0500

On Wed, 2014-11-05 at 18:18 +0000, Sorensen, Jesper wrote:
> Hi Jason,
> I’d like to follow up on this… Is there a way in post-processing to
> recalculate the pressure?

Sure. Do a single-point calculation for each of the frames in your
trajectory. I think if you request an mden file, it will print out the
pressure there.

Otherwise, you can request a 1-step NPT calculation for each frame,
which will print out the pressure (unless you're using the MC barostat
in pmemd, obviously).

> I set the pressure with the Berendsen barostat first before switching
> to MC, so I imagine that it’s where I’d like it to be on average (I
> realize it fluctuates a lot), but it’d still be nice to check that it
> hasn’t diverged too much due to unforeseen events…

I still maintain that the density is a better property to track than the
pressure. That was one of my tests I used to validate the MC barostat
> Is it not possible to asynchronously compute the virial for the energy
> output using pmemd.cuda maybe by throwing it to the CPU similar to
> writing files? This way the simulation can proceed, keeping it’s
> efficient speed, yet outputting this property similar to the Berendsen
> barostat… I guess this would also be a good question for Ross Walker
> since it involves the GPU version of pmemd.

I can almost guarantee that this won't happen, as the work involved is
considerable. As I understand it, launching GPU kernels is a blocking
action (that is, you _must_ wait for the kernel to finish/return before
progressing). So you would have to start a new thread to do this, which
opens up a whole new can of worms (I don't think any of the major
routines in pmemd are thread-safe, for instance). Not to mention that a
lot of changes would be necessary to even permit energy calculations to
be performed on the CPU in pmemd.cuda in the first place.

In summary, it's possible "in principle", but impractical in pmemd.cuda.
Developer (i.e., Scott's ;) time would be better spent on other
improvements when working on pmemd.cuda.


Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Wed Nov 05 2014 - 11:00:03 PST
Custom Search