GROMACS

Removing particle decomposition requires that the regression test perl script stops always triggering -pd when simple grid searching is being used. Currently, the complex/sw test uses the shell-model polarization code, with simple, thus -pd (and vsites and "flexible" constraints). This would have to change to grid+DD when we remove PD.

Using installs of either master or release-4-6 HEAD with

gmxtest.pl complex -only sw\\b

by changing the grompp.mdp and/or gmxtest.pl files, I observed:
-pd and simple passes the test with 1-8 tMPI ranks
-pd and grid passes the test with 1-8 tMPI ranks
-dd and simple passes the test with 1 rank, and gives the expected fatal error with DD+simple unsupported (perl script hack required to remove the hard-coded -pd flag)
-dd and grid passes the test with 1 rank, and fails with all higher numbers

The fails are pretty subtle:

comparing energy file reference_s.edr and ener.edr

There are 31 terms in the energy files

There are 5 terms to compare in the energy files

Angle            step  19:       71.2642,  step  19:      71.1485
Angle            step  20:       70.2909,  step  20:      70.1638

Files read successfully

So, if there's a problem, then I think it is already present in the code and release-4-6 should get a fatal error for shells + DD and ranks > 1. It could have the same cause as the problem that was swept under the carpet in #993 (the forces may not be very reproducible in single precision), but when I tested release-4-6 in double precision I got the same error with DD, grid, and -ntmpi > 1.

To see if OpenMP was useful as parallelism, I tried to test the shell code with the Verlet scheme (using verlet-buffer-drift=-1 because the test case has a box that's only just large enough for the unmodified rlist), and it segfaulted with both grid and simple, even with just 1 rank and 1 thread. So I believe that needs a fix or a fatal error, also.

The above poses a problem for the proposed removal of PD - the existence of what might be an historical bug reduces the support in 5.0 for shells from "MPI with group scheme and PD" to "serial with group scheme." We have a few open bugs for / suspicions of the shell code (#879, #993, #713). I suggest we proceed with removing PD, and require the complex/sw test to run with 1 rank, unless/until someone wants to find out what the real problem is. Hopefully, with the interest in polarizable models expressed in #1292 that will be sooner rather than later!