Re: AMBER: Scaling for octahedral versus rectilinear boxes

From: Robert Duke <rduke.email.unc.edu>
Date: Fri, 31 Aug 2007 08:40:48 -0400

Hi David -
I think the major expense is that one needs to use all the unit cell vectors
when calculating imaged coordinates (ie., deriving the primary unit cell
coordinates for an atom from it's "real" coordinates - the coordinates
currently associated with the atom). This occurs in some inner loops and
junks the code up a bit, mostly for the sake of processors that don't do a
good job of branch prediction (so on an intel ia32 chip, branch prediction
is really good for something like this and the conditional really costs next
to nothing, so while I duplicate some chunks of code to avoid conditionals,
this does not matter that much for good branch predictors). So I have never
done scaling tests per se for an orthogonal box vs. a truncated octahedron;
it probably depends mostly on total atom count (assumed to be lower for the
same "problem" solvated in a truncated octahedron vs. an orthogonal unit
cell) and on how the fft gridding works out. Generally larger problems
scale better due to better spatial decomposition (there is more space to
decompose, with less overlap), which just says that you can productively run
a larger problem on more processors. This is one reason that people trying
to show high scaling pick the biggest problem they can, regardless of
whether one can currently get adequate statistics for these really large
problems.
Regards - Bob

----- Original Message -----
From: "David Cerutti" <dcerutti.mccammon.ucsd.edu>
To: <amber.scripps.edu>
Sent: Friday, August 31, 2007 1:32 AM
Subject: AMBER: Scaling for octahedral versus rectilinear boxes


> Hi,
>
> I think I know who can answer this, but I'm going to the listserv so the
> knowledge can be spread around.
>
> Do simulations with truncated octahedral boxes scale as well to many
> processors as those with rectilinear boxes?
>
> On the one hand, the neighbor list may level the playing field, but on the
> other, there may still be some need for a domain-decomposition style
> computation to generate that list, and the best thing to have then would
> be an orthonormal box.
>
> Thanks!
>
> Dave
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>


-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Sep 02 2007 - 06:07:30 PDT
Custom Search