Re: amber-developers: RGBMAX=25A by default?

From: David A. Case <case.scripps.edu>
Date: Thu, 21 Apr 2005 18:18:53 -0700

On Thu, Apr 21, 2005, Yong Duan wrote:
>
> The default RGBMAX in AMBER 8 is 25A which actually makes GB slower than
> most explicit PME simulations. Can some realistic numbers be set as the
> default instead? Of course, this performance justifies the fully
> polarizable simulations pretty well :).
>

A few comments:

1. I think the original poster was comparing PB_MD (with a shell of
explicit
water, i.e. igb=10) with GB (igb=2). [It was ambiguous, since "PB" could
mean
Poisson-Boltzmann or periodic boundaries, but I think it was the former.]
Do
other people have any similar timing comparisons (or "quality"
comparisons,
for that matter)? Please post to the Amber wiki.

2. GB is changing so fast that few people have had time to really try to
optimize the codes for speed; on the other hand, PME is very mature, and
Bob
Duke and others have spent a lot of time getting fast code there. Alexey
and
John M. have some pretty promising new GB developments, which may change
people's thinking. The bottom line is (in my view): we should still be
concentrating on getting good quality GB runs: speed will come later.

Especially right now, when our ideas of what constitutes the optimal GB
model
are in flux, I encourage people to be cautious about spending lots of time
trying to speed up GB. Also, there are some very good ideas in NAB that I
think should be ported over. But I welcome eager volunteers, as long as
they
keep in close contact so that speed and quality development go
hand-in-hand.

3. One can fiddle with rgbmax (and with cut, nrespa and nrespai) to
optimize
perfomance, and this should indeed be done. A good starting point is
cut=rgbmax=15, nrespa=4, nrespai=1 or 2. But the fundamental problem is
that
we are not using any fast method for long-range interactions, and this is
inevitably will be very bad for bigger systems. [A pair list is also
needed,
but really only pushes back a little bit the onset of poor scaling.]

4. Don't forget that GB explores conformational space a lot faster than
explicit solvent simulations do. The difference can be an order of
magnitude,
and should be taken into consideration for some types of simualations.

....cheers....dac
Received on Wed Apr 05 2006 - 23:49:57 PDT
Custom Search