Re: [AMBER] Input case causing NaN in sander (and possibly pmemd) (Ambertools17)

From: Åke Sandgren <ake.sandgren.hpc2n.umu.se>
Date: Mon, 29 May 2017 22:18:48 +0200

I can't see any obvious problems using checkoverlap, but on the other
hand i'm not quite sure if i'm using it correctly.

The input files are available from

https://web-upload.its.umu.se/dl/f/1496089012_27796/inf-nan-problem.tar.bz2

for 3 days. If you haven't grabbed them by then i could send it in mail,
it's only 1.1M.

On 05/29/2017 09:43 PM, David Case wrote:
> On Fri, May 19, 2017, Åke Sandgren wrote:
>>
>> One of our users have a input set that is generating "inf" in
>> the call vdinvsqrt(icount, cache_r2, cache_df) in short_ene line 392
>
> This sounds like a case where two atoms are on top of each other:
>
>> delr2 ends up as 0 which is of course < filter_cut2
>> This is then propagated to a NaN in the
>
>> Eventually this problem propagates down to fill_charge_grid where
>> fr1/2/3 are all NaN, causing k0, j00, i00 to become 2147483645, or there
>> about, and finally i0 is used as index into q causing a segfault.
>
> I agree that we should not get a segfault here, but would like to know
> more about the actual reason that delr2 is zero.
>
> Could you ask the user to use the cpptraj "checkoverlap" command to see
> if there are close atomic contacts? That would help in narrowing down
> the problem.
>>
>> I wonder if the delr2 < filter_cut2 in ew_directp.h should check that
>> delr2 != 0 too?
>
> This could be done, but we would have to see if it would have an impact
> on performance -- I'm wary of putting extra if statements into inner loops.
> (And while I agree in principle that programs should not segfault, there
> are places in high-performance code where the (simple) cure may be worse
> than the disease.
>
> If the user doesn't have an obviously bad structure, see if you can send
> files that would allow us to try to reproduce the problem.
>
> Thanks for the report....dac
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>

-- 
Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden
Internet: ake.hpc2n.umu.se   Phone: +46 90 7866134 Fax: +46 90-580 14
Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon May 29 2017 - 13:30:03 PDT
Custom Search