[gmx-users] PME/Ewald

[Jason de Joannis]

> Date: Tue, 29 Mar 2005 16:19:10 +0100 (BST)
> From: "M.Naser" <mn2 at hw.ac.uk>
> Subject: [gmx-users] PME/Ewald
> To: "Discussion list for GROMACS users" <gmx-users at gromacs.org>
> Message-ID: <2195.137.195.36.211.1112109550.squirrel at wmail3.hw.ac.uk>
> Content-Type: text/plain;charset=iso-8859-1
>
> Hi All User,
>
> Is there any standerd system size beyond which I should use PME (not
> Ewald) to aviod overhead in setting up grids and transform?
>

 In principle there is a break-even point because of the overhead
involved with PME. The PME method scales as NlogN compared with
Ewald as N^3/2 so PME prevails as you increase the number of charges N.
In practice the actual value of N where performance is equal is
 a) Small eg. <100
 b) Hard to generalize since it depends on many things including
    the accuracy required, the computer's architecture and the
    FFT code.

 Moreover, the Ewald implemented in Gromacs will have a much lower
or nonexistent break-even point because it is neither parallelized
nor is it optimized for the selection of Ewald parameter (and therefore
scales as N^2).

 Long story short, use PME even with only two charges, unless you
don't care about performance. Probably the reason why Ewald is included
is for debugging the FFT-MPI stuff.

 /Jason

-- 
Jason de Joannis, Ph.D.
Emory University
1515 Pierce Dr, Atlanta, GA 30329
Phone: (404) 712-2983
Email: jdejoan at emory.edu
http://userwww.service.emory.edu/~jdejoan