Re: [AMBER-Developers] Ascii Restart Files: precision on velocities

From: Jason Swails <>
Date: Mon, 7 Nov 2011 17:33:21 -0500

On Mon, Nov 7, 2011 at 5:24 PM, <> wrote:

> I finally beat some sense into the way that mdgx is starting and stopping,
> what it calls the first step, etc. Hadn't been too worried about that
> when I was more interested in getting long stable runs out of it, or when
> I was just accumulating everything I needed internally.
> But it's alerted me to one issue, which I had foreseen but not thought
> much about. The ascii restart format is much less precise on velocities
> than it could be. We should be able to squeeze another factor of 20.45 in
> there by encoding velocities at their internal units of A/ps rather than
> scaled down by sqrt(418.4). It's really the velocities that limit the
> continuity of restarts.
> I suppose it's not so much an issue if we're moving towards standardized
> binary restarts anyway, but how did we come up with a format that is both
> cryptic and less precise than the %12.7f format can handle? I realize
> that, in some case, the velocities could be any values because of some
> very large random value, but the probability of getting an absolute
> |velocity| > 999 A/ps has to be so remote, even at like 1500K, that it
> shouldn't break the format in a lifetime of number crunching. Not meaning
> to start a flame war, but just curious on this one.

I'm guessing that the effort involved in changing the standard safely at
this point (and the risk of breaking backwards-compatibility for both ours
and 3rd-party programs anyway) isn't worth the gain, especially with the
work Dan's done to implement a NetCDF version.

I'll accept your assertions that the velocities are the biggest issues, but
the only thing that tweaking this format does for us at this point is get
us a little *closer* to a true restart, but still nowhere close. Even if
the velocities have less precision than they could, I think they're
probably sufficient for restarting a simulation (in that we don't have to
re-equilibrate). Maybe not for uber-sensitive NVE simulations with a tiny
time step, but I don't know that anything less than our NetCDF restart will
satisfy there.

My 2c,

Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
AMBER-Developers mailing list
Received on Mon Nov 07 2011 - 15:00:02 PST
Custom Search