Project

General

Profile

Bug #313

gmxtest-4.0.4 fails for several kernel tests; likely due to LJ14 interaction calculations

Added by Dmitry Kopelevich over 10 years ago. Updated over 10 years ago.

Status:
Closed
Priority:
Normal
Assignee:
Erik Lindahl
Category:
mdrun
Target version:
Affected version - extra info:
Affected version:
Difficulty:
uncategorized
Close

Description

Created an attachment (id=360)
File produced by ./gmxtest.pl kernel for kernel020

I've run gmxtest-4.0.4 for gromacs-4.0.4 on x86 (i32) and x86_64 Linux systems.

The following kernel tests reported errors in the virial and potential energy tests:

kernel020 (2 errors in checkpot; 7 errors in checkvir)
kernel120 (2 errors in checkpot; 7 errors in checkvir)
kernel121 (2 errors in checkpot; 7 errors in checkvir)
kernel122 (1 error in checkpot; 7 errors in checkvir)
kernel123 (2 errors in checkpot; 9 errors in checkvir)
kernel124 (1 error in checkpot; 7 errors in checkvir)
kernel220 (2 errors in checkpot; 7 errors in checkvir)
kernel221 (2 errors in checkpot; 7 errors in checkvir)
kernel222 (1 error in checkpot; 7 errors in checkvir)
kernel223 (2 errors in checkpot; 9 errors in checkvir)
kernel224 (1 errors in checkpot; 7 errors in checkvir)
kernel320 (2 errors in checkpot; 7 errors in checkvir)
kernel321 (2 errors in checkpot; 7 errors in checkvir)
kernel322 (1 error in checkpot; 4 errors in checkvir)
kernel323 (2 errors in checkpot; 9 errors in checkvir)
kernel324 (1 errors in checkpot; 4 errors in checkvir)

The numbers of the reported errors are the same for both single and double precision on both the 32-bit and 64-bit systems. Also, for each of the failed tests, the energy values reported in files chekpot.out are essentially independent of the system architecture and precision.

All energy errors appear to be related to the calculation of the LJ14 potential energy; in all of the failed tests, the LJ14 energies are shifted by the same amount with respect to the reference values.

As an example, I am attaching the checkvir.out file produced by gmxtest.pl on a x86_64 system for a single precision test kernel020.

For comparison, gmxtest did not report any errors for kernel tests for gromacs-4.0.3. There were also some small errors in complex tests for both of the 4.0.3 and 4.0.4 gromacs versions but these errors could be attributed to numerical accuracy.

checkvir.out (1.77 KB) checkvir.out File produced by ./gmxtest.pl kernel for kernel020 Dmitry Kopelevich, 04/08/2009 06:22 PM

History

#1 Updated by Mark Abraham over 10 years ago

This bug has been linked from the mailing list. On my first read-through I didn't notice that it was specific to Buckingham interactions, and made a mental note to check its consequences for me. Having noticed, I'm actually OK. I thought that a post here reflecting that there's no sign of non-Buckingham problems might avoid many people panicking unduly.

#2 Updated by Erik Lindahl over 10 years ago

This turned out to be an old bug in Gromacs-3.3 that we used to generate the reference files, while Gromacs-4.0 actually has it right.

I think this is very unlikely to have affected any real simulations. It only occured in my simple test system where I was using Buckingham normal nonbonded interactions, but Lennard-Jones 1,4 interactions (don't ask why...). pdb2gmx will never generate any Buckingham interactions, and in the few cases where people use them I think it is unlikely that it would have been combined with standard Lennard-Jones interactions (but never say never).

In any case, the fix was committed in the release-3-3-patches branch earlier today, and I have also updated+committed new reference files for the gmxtest module so all kernels pass all tests formally too (together with a bunch of other fixes from Mark Abraham).

Cheers,

Erik

Also available in: Atom PDF