Project

General

Profile

Bug #1308

When running normal modes on a structure, unless one thread is used, the energy will not be the same as the previous minimization

Added by Michael Shirts about 4 years ago. Updated almost 4 years ago.

Status:
Closed
Priority:
Normal
Assignee:
Berk Hess
Category:
-
Target version:
4.6.4
Affected version - extra info:
Affected version:
4.6
Difficulty:
uncategorized
Close

Description

As noted above:

No idea who to assign to.

When I minimize a structure, I can get down to the force max being <0.01

Low-Memory BFGS Minimizer converged to Fmax < 0.01 in 6839 steps
Potential Energy = -5.12340607768673e+03
Maximum force = 6.68907856457542e-03 on atom 3029
Norm of force = 2.19978176343026e-03

kJ/nm. However, when I try to perform a normal mode analysis, it
complains that the maximum energy is >30 kJ/mn.

Non-cutoff electrostatics used, forcing full Hessian format.
Allocating Hessian memory...

Maximum force: 3.91984e+01
Maximum force probably not small enough to ensure that you are in an
energy well. Be aware that negative eigenvalues may occur when the
resulting matrix is diagonalized.

Fix was to run with -nt 1, and then the maximum force was exactly the same.

test_benzene_min.sh View (942 Bytes) Michael Shirts, 08/02/2013 04:24 AM

benzene_solid_minfornm.mdp (3.93 KB) Michael Shirts, 08/02/2013 04:24 AM

benzene_solid_nm.mdp (3.63 KB) Michael Shirts, 08/02/2013 04:25 AM

benzene_256.top (110 Bytes) Michael Shirts, 08/02/2013 04:25 AM

benzene_p2_256_minfornm.gro (135 KB) Michael Shirts, 08/02/2013 04:26 AM

benzene.itp (3.88 KB) David van der Spoel, 08/12/2013 10:45 AM

Associated revisions

Revision 7b5c0b83 (diff)
Added by Berk Hess about 4 years ago

fixed parallel normal modes with PME

When doing a normal mode calculation with PME and (thread-)MPI,
the PME forces were incorrect. Also fixed some EM/NM output layout.
Fixes #1308

Change-Id: Ia7862fa62e235336c546824afdcbe28f37b145c5

History

#1 Updated by Mark Abraham about 4 years ago

Thanks.

I imagine only Berk or Erik have any real idea how this works. I'll force some appropriate usage for 4.6.4 if there's no fix planned.

#2 Updated by David van der Spoel about 4 years ago

Is there a test set? I assume this is not done using the .gro output file from the minimization?

#3 Updated by Mark Abraham about 4 years ago

David van der Spoel wrote:

Is there a test set?

Ahahahaha :-)

(I have master-branch machinery to add lightweight versions of such tests mostly ready, but not yet fit for routine use. Some of Teemu's more recent stuff with options parsing will be very useful.)

I assume this is not done using the .gro output file from the minimization?

#4 Updated by David van der Spoel about 4 years ago

I meant: is there an example that shows this problem? It would be easier to work with Michael's example than cough up something myself.

#6 Updated by Michael Shirts about 4 years ago

Test case uploaded. Run the script with changing the number of threads. When the number of threads is 1, the force in the normal modes calc equals the force after minimization. When not, it doesn't. Restart is done with .trr.

#7 Updated by David van der Spoel about 4 years ago

Hi Michael,
benzene.itp is missing...

#8 Updated by David van der Spoel about 4 years ago

OK nevermind, I fixed a itp file myself and have reproduced the problem.

#9 Updated by David van der Spoel about 4 years ago

Potential fix uploaded: https://gerrit.gromacs.org/#/c/2549/ but not tested more than for reproducing the EM energy (which it does now). There may still be other reasons why parallel NM does not work whatsoever.

#10 Updated by Berk Hess about 4 years ago

Can someone attach a OPLS benzene.itp?
I tested with my own methanol system and I can't reproduce this issue.

#11 Updated by David van der Spoel about 4 years ago

Attached.

#12 Updated by Berk Hess about 4 years ago

Never mind.
I used Michael's mdp file and reproduced the issue. The problem is with PME. I assume some PME component is added multiple times.

#13 Updated by Berk Hess about 4 years ago

  • Status changed from New to Fix uploaded
  • Assignee changed from Mark Abraham to Berk Hess

NM parallelizes over gradient components (no particle/domain decomposition). This means every MPI rank does a separate force calculation.
However, PME was still initialized (on every rank) with a parallel setup.
I assume 4.5 was also affected by this. I uploaded a proper fix.

#14 Updated by Berk Hess about 4 years ago

  • Status changed from Fix uploaded to Resolved
  • % Done changed from 0 to 100

#15 Updated by Mark Abraham almost 4 years ago

  • Status changed from Resolved to Closed

Also available in: Atom PDF