Updated benchmark script

[peastman]

This updates the benchmarking script to use LangevinMiddleIntegrator and a larger time step where possible. Here are the speeds I get for the PME benchmark with different possible combinations of settings.

LangevinIntegrator, 2 fs, HBonds constraints: 417
LangevinMiddleIntegrator, 4 fs, HBonds constraints: 740
LangevinMiddleIntegrator, 5 fs, AllBonds constraints, HMR: 568
Langevintegrator, 5 fs, AllBonds constraints, HMR: 775

The ordering is the same for most benchmarks. Using LangevinMiddleIntegrator with the larger step size produces a big speed improvement compared to LangevinIntegrator. But the very fastest combination is to use LangevinIntegrator with HMR and more constraints. Since LangevinMiddleIntegrator needs to apply constraints more often, it takes a big hit from using AllBonds constraints.

Therefore if you specify --heavy-hydrogens, I made the script continue to use LangevinIntegrator in place of LangevinMiddleIntegrator.

[jchodera]

Can we deprecate the AllBonds constraints? This can significantly modify the effective torsions, and it doesn't seem to give any significant performance advantage over LangevinMiddleIntegrator + 4 fs + HBonds.

[jchodera]

cc our previous discussion here: #2520

[jchodera]

Presumably, there should also be an MTS Langevin scheme that should give performance just as good as any AllBonds combination but without the pathologies associated with it.

[peastman]

I wouldn't deprecate that option. It works fine and does exactly what it says. Whether it's appropriate for a particular force field is a different matter.

Experimenting with different MTS protocols could be interesting. We could probably find something that allows a 5 fs outer time step.

[jchodera]

I wouldn't deprecate that option.

We shouldn't remove the AllBonds option. I'm just saying we should remove it from the benchmark options for Amber force fields, since its inclusion comes with our tacit approval that this is a sensible combination. It is not.

It works fine and does exactly what it says. Whether it's appropriate for a particular force field is a different matter.

Are there force fields for which this is an appropriate combination? If so, can we use that force field instead?

Again, I think we want OpenMM to "do the right thing" by default. If one of our standard benchmarks does not do the right thing, why do we want this?

[jchodera]

Another alternative to all-bonds would be to use the now-popular virtual sites approach to representing hydrogens that gromacs has popularized. There are at least publications about this:
https://onlinelibrary.wiley.com/doi/full/10.1002/jcc.26198?af=R

[peastman]

I'm fine with not including those on our benchmarks page.

Are there force fields for which this is an appropriate combination?

I think it's pretty common in coarse grained models.

Another alternative to all-bonds would be to use the now-popular virtual sites approach to representing hydrogens that gromacs has popularized.

Using vsites for hydrogen is in addition to constraining all bonds, not instead of it. Doing that removes even more degrees of freedom than just constraining bonds.