[gmx-users] "nstlist=-1" and performance problem

[Berk Hess]

Hi,

But grompp should have given a warning about this.
You should have only obtained a tpr file when using the -maxwarn option of grompp.
Or did grompp not give a warning about the buffer being too small?

Berk

From: lulanyuan at msn.com
To: gmx-users at gromacs.org
Subject: RE: [gmx-users] "nstlist=-1" and performance problem
Date: Wed, 28 Jan 2009 17:12:00 -0500








Hi Berk,
Thanks very much. I've found the problem. As you said I forgot to increase rvdw. So the buffer is zero and the system needs to update the nblist every step, which is obviously wrong. After I corrected the mistake, the speed for "nstlist=-1" is roughly three times faster as I expected.
Lanyuan

From: gmx3 at hotmail.com
To: gmx-users at gromacs.org
Subject: RE: [gmx-users] "nstlist=-1" and performance problem
Date: Wed, 28 Jan 2009 10:42:59 +0100








Hi,

I assume you did not only change nstlist, but also other mdp settings, or not?

Changing only nstlist from 1 to -1 should result in nearly equal or much faster speeds.

For efficient simulations with nslist=-1 you need to set an appropriate
buffer region (rlist > rcoulomb=rvdw). In 4.0 you have to optimize this by hand.
It would be nice to have this automated.
At the end of the log file (a run of a few minutes is enough) you can see what
the neighborlist lifetime is. Optimal is usually something between 10 and 20.
Play with rlist to optimize the performance (also printed at the end of the log file).

Berk

From: lulanyuan at msn.com
To: gmx-users at gromacs.org
Date: Mon, 26 Jan 2009 17:31:43 -0500
Subject: [gmx-users] "nstlist=-1" and performance problem








Hello,
Could Berk or someone else answer my question regarding the new "nstlist=-1" option in gmx4? In my understanding, the algorithm for this option is the one used in DLPOLY/LAMMPS. However, When I was trying two gromacs runs with "nstlist=-1" and "nstlist=1", I found the speeds were basically the same. Isn't this automatic nblist checking supposed to be faster than updating the nblist every step? I also did similar tests on LAMMPS for the same system and found the speed for automatical updating was several times faster. As a result, although Gromacs is about 60% faster if we choose nstlist=1,  it's much slower than LAMMPS if we choose the automatic nblist updating. 
Thanks in advance for any ideas about it.
Lanyuan Lu


不加好友也能聊？试试MSN在线通！ 现在就下载！
Express yourself instantly with MSN Messenger! MSN Messenger
八卦娱乐包打听，MSN资讯速递帮你忙！ 了解详细！
_________________________________________________________________
See all the ways you can stay connected to friends and family
http://www.microsoft.com/windows/windowslive/default.aspx
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20090128/8dd04240/attachment.html>