This error message "an illegal memory access was encountered launching
kernel kNLSkinTest" is not one of the code's most helpful error messages.
It's kind of like saying that the patient died from lack of oxygen reaching
the brain. (But WHY--what started the sequence of events leading to that
condition?) In the GPU code, if something goes wrong, the ultimate symptom
is often that a particle gets moving too fast and flies off into the
stratosphere, which creates the kNLSkinTest problem. But if you can post
your inputs we can take a look. For some very large systems we have been
seeing occasional issues like this in standard MD and I have made some
progress tracking down the problem.
Dave
On Sat, Jun 23, 2018 at 10:53 AM Maximilian Ebert <mebert.chemcomp.com>
wrote:
> Hi,
>
> we are running the following TI protocol using AMBER18 pmemd.cuda:
>
> Full dual topology one step TI transformation. 2fs time step and full
> shake.
>
> Common equilibration at lambda 0.5 and scalpha = scbeta = 0
> 1. minimization on cpu
> 2. heating to 300k 200ps
> 3. nvt 300k 200ps
> 4. npt (Monte Carlo) 1bar 300k 1600ps
>
> TI with 14 lambda windows (values from chebyshev integration) and
> scalpha = 0.5 scbeta=12
> 1. start with new velocities from common equilibration NPT 1000ps
> 2. production NPT 1000ps
>
> each trajectory is segmented in a max length of 250ps and restarted.
>
> The above protocol works without any crash for small ligands like the
> lysozyme + bnz/phn but using larger systems like kinase + Imatinib
> (gleevec) causes for edge lambdas (lambda > 0.8 or lambda < 0.2) the
> system to crash.
>
> Nothing in the stdout indicates a reason for the Error: an illegal
> memory access was encountered launching kernel kNLSkinTest. Also
> restarting the segment often recovers and runs just fine. The message
> comes most of the times in the beginning of the run after a couple of
> steps but not always. The Skin appears to have something to do with
> the barostat and the box size.
>
> Therefore, we also tried the berendsen barostat but this resulted
> mainly in distorted ligand conformations (forming a pile of atoms in
> space). There might also be a bug in the berendsen barostat when using
> TI GPU code. We also tried 1fs with tishakemask for the ligand but
> didn't help.
>
> We are currently running a test with full NVT in each lambda
> production window. However, if anybody can give us some insight what
> is going on it would help us greatly.
>
> Thanks,
> Max
>
> --
> Maximilian Ebert, PhD | Applications Scientist
> +1 (514) 393-1055 x136 | support.chemcomp.com
>
> --
> Chemical Computing Group | www.chemcomp.com <http://www.chemcomp.com> |
> 910-1010 Sherbrooke W., Montreal QC H3A 2R7, Canada | +1 514 393 1055
>
> Manage Communications: www.chemcomp.com/Manage_Subscription.htm
> <http://www.chemcomp.com/Manage_Subscription.htm>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Jun 23 2018 - 15:30:02 PDT
