From: Ajasja Ljubetič (ajasja.ljubetic_at_gmail.com)
Date: Wed Sep 10 2014 - 19:58:48 CDT
On 11 September 2014 02:51, Tristan Croll <tristan.croll_at_qut.edu.au> wrote:
> ... and now I see I'm just running over old ground. Sorry – I did do a
> search, but obviously didn't hit the right keywords. For anyone that
> comes across this thread in future, the problem (as hashed out in detail
> here:
> http://www.ks.uiuc.edu/Research/namd/mailing_list/namd-l.2013-2014/3007.html)
> is that the colvars calculations are running on a single core).
>
> This is also described in the user guide
<http://www.ks.uiuc.edu/Research/namd/2.9/ug/node55.html#SECTION000132410000000000000>
:
- *Performance issues:* While NAMD spreads the calculation of most
interaction terms over many computational nodes, the colvars calculation is
not parallelized. This has two consequences: additional load on the master
node, where the colvar calculation is performed, and additional
communication between nodes. NAMD's latency-tolerant design and dynamic
load balancing alleviate these factors; still, under some circumstances,
significant performance impact may be observed, especially in the form of
poor parallel scaling. To mitigate this, as a general guideline, the size
of atom groups involved in colvar components should be kept small unless
necessary to capture the relevant degrees of freedom.
>
This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:21:13 CST