From: JC Gumbart (gumbart_at_physics.gatech.edu)
Date: Wed Sep 27 2017 - 12:55:49 CDT
Yes, the one released in Dec. I haven’t tested the latest build yet; I’m just using what’s available on Bridges at the moment.
I agree it seems to be related to the new kernel, although the fact that it’s strongly GPU dependent is really weird to me.
I’ll send the files shortly.
> On Sep 27, 2017, at 1:40 PM, Hardy, David J <dhardy_at_illinois.edu> wrote:
> Hi JC,
> When you say "2.12" you do mean the version of NAMD released back last December, yes? There have been some other CUDA development and bug fixes since then. Have you tested any more recent versions of NAMD, say any nightly builds since August?
> Your results for versions 2.10 and 2.11 demonstrate that the issue you see doesn't affect the older CUDA kernels, just the new ones.
> Would you be willing to, off-list, send me your data files so that I can test this for myself?
> David J. Hardy, Ph.D.
> Theoretical and Computational Biophysics
> Beckman Institute, University of Illinois
>> On Sep 27, 2017, at 11:48 AM, JC Gumbart <gumbart_at_physics.gatech.edu> wrote:
>> I know this has been discussed at length previously, but I may have misunderstood something. I have a system that’s pretty small but the patch grid is still always at least 2 in any dimension. I first run a minimization/equilibration on a p100 with NAMD 2.12 with “margin 5”; the size in z stabilizes at ~48 A. I then run a continuation with no margin. The size in z quickly drops to ~42 A (within 5000 steps). This is still on the same GPU. If I turn “margin 5” on again, it stays around 48 A.
>> I’ve tested numerous combinations; it seems the CPU version always drops in size regardless of margin as do previous GPU versions. But what’s interesting is that in 2.12 it seems to depend on the specific GPU.
>> all cpu versions: 48 A -> 42 A
>> 2.10 on k20c, 2.11 on p100: 48 -> 42
>> 2.12 gpu on k20c: 48 -> 42
>> 2.12 gpu on p100: 48 remains constant
>> 2.12 gpu on k80: 48 -> 90 in 5k steps (????)
>> Regardless of version, if margin is less than 5, it drops. Note that for a couple hundred other runs of very similar systems before (same size within a few Angstroms in each dimension), I didn’t see this behavior. It appears to be very specific to the precise size of this one.
>> Any thoughts? Dave noted before that there should be other instabilities, but with the exception of the behavior on k80s, the change in size by 15% seems to be the only effect.
This archive was generated by hypermail 2.1.6 : Sun Dec 31 2017 - 23:21:41 CST