Date: Wed, 4 Dec 2002 08:28:43 -0800
From: "David A. Case"
Subject: Re: gB vs. explicit

On Tue, 3 Dec 2002, Kristina Furse wrote:

> I've been running sander calculations on a large protein (cyclooxygenase
> dimer with 2 heme's, 2 arachidonic acids and 8 bog detergents--1118
> residues, 18,422 s). In the interest of saving compute cycles, I've been
> using generalized Born and keeping the 200 crystallographic waters held in
> place with 1kcal positional restraints.

> I recently set up an explicit water simulation as a control (truncated
> octahedron, 11A buffer, 6 Na+, SPC/E water...): Scary big (29,726 residues,
> 104,234 atoms)--so I assumed this calculation would take forever. But I've
> found that the explicit calculation is running nearly twice as fast as the
> generalized Born version! Have you guys really improved explicit
> calculations this much???

You might ask instead: why is the GB so slow??? :-)

With the GB system and a 12 Ang cutoff, you are calculating a lot of fairly
expensive interactions, and you also have to compute the GB radii and their
derivatives.

With the explicit solvent, you only need to do interactions out to 8 or 9
Angstroms, with the reciprocal part doing the rest. (There is not yet any
"reciprocal" analogue for GB.) The down-side is that (in parallel runs) you
have to communicate all the water positions between processors.

Up until now, we had never heard from anyone trying to use GB on an 18,000
atom system, and it is clear that such a problem needs some attention in order
to obtain better performance. (For the most part, we have concentrated here
on checking and improving the fidelity of the GB model, not on speeding it
up.) We have your GB input files, and will be looking at performance on that
system as well develop the next version of the codes.

...thanks for your comments....dac