RE: ABF (NAMD 2.7x) speed ?

From: Ziemys, Arturas (Arturas.Ziemys_at_uth.tmc.edu)
Date: Wed May 26 2010 - 07:24:54 CDT

Hi, Jerome,

Thank you a lot for the insight. I was following tutorial explicitly before too blindly... However, when I have declared just one atom in ref group the performance increased x10 times.

Best
Arturas

________________________________________
From: heninj_at_gmail.com [heninj_at_gmail.com] On Behalf Of Jérôme Hénin [jhenin_at_ifr88.cnrs-mrs.fr]
Sent: Wednesday, May 26, 2010 4:52 AM
To: Ziemys, Arturas
Cc: namd-l
Subject: Re: namd-l: ABF (NAMD 2.7x) speed ?

Hi Arturas,
[copying this to the mailing list again]

Your "old ABF" and colvars script are not equivalent. In the "old ABF"
script, the reference group abf1 contains a single atom, while in the
colvars script, it is much larger:

colvars: Initializing atom group "ref".
colvars: Atom group "ref" defined, 4432 initialized: total mass = 88831.8.

An atom group containing thousands of atoms is going to be very
detrimental to scaling, because of the additional inter-process
communication involved. There is a warning about this in the user's
guide, but it is not very clear. I'll try to make it clearer and
stronger.

In this case, you could try and find a smaller set of atoms which has
approximately the same center of mass. Tens of atoms (or maybe a few
hundred) can give reasonable results. Then you can benchmark the old
ABF code using the same atom groups, and tell us about it.

Cheers,
Jerome

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:54:10 CST