[gmx-users] Non-linear slowdown in grompp with very many bonded interactions

[Chris Neale]

Hello,

I am trying to figure out why grompp is so slow when given a long lists of
bonded interactions. For example, my base system takes 7 seconds to run
grompp and 37 seconds to do 100 steps of EM. However, when I add ~250,000
distance restraints grompp takes 3 minutes, and when I add ~600,000
distance restraints grompp takes 18 minutes. In both cases, the mdrun
command is almost as fast as it was without any distance restraints (~42
seconds in each case). I made a bit of progress by converting the distance
restraints to type-6 bonds, in which case the addition of ~250,000 or
~600,000 type-6 bonds results in grompp finishing in 1.5 minutes and 6
minutes, respectively.

These numbers indicate to me that there is some type of non-linear slowdown
when adding bonded interactions (like possibly a check for duplication of
bonded terms as each one is read in, or memory reallocation).

I would appreciate it if anybody knows which routine I should look at to
see if I can get rid of this grompp lag. I am currently working with
gromacs 5.1.2. Newer versions should be OK, but since ideally I want to use
energy group exclusions, gromacs-2020 and it's verlet-scheme-only code is
probably not an option for me. However, I did verify that I get the same
type of non-linear slowdown when adding many bonded terms to gromacs-2020.

Thank you,
Chris.