From: Harish Vashisth (harish.vashisth_at_gmail.com)
Date: Tue Jul 22 2014 - 13:00:04 CDT
Hi Jerome, Aron:
Thanks for your suggestions. I did a quick test. Previously, we were using
the width parameter of 5 for sampling between 0 and 180 of each dihedral
angle CV. I experimented with increasing this to 10 and 20, which allows
usage of more (read 8) CVs (Not that it is scientifically meaningful or
will give useful information). So as Jerome suggests memory issue is
sensitive to binwidth which is coupled with large number of CVs as each CV
is to be sampled.
Aron: i double checked that our definitions of CVs are correctly written; I
think the error was not due to wrong syntax. Also, i appreciate the issue
with sampling many CVs, but thought something on the order of 20 may be
possible with abf. We were experimenting with running abf calculation using
backbone dihedrals of a 12-15 residue peptide. we were also using a single
ABF bias on all CVs in this test case like below:
colvars phi1 psi1 phi2 psi2 phi3 psi3 phi4 psi4
We have also tried using "metadynamics" as a bias in NAMD with similar
memory error message with large number of CVs.
On Tue, Jul 22, 2014 at 1:32 PM, Aron Broom <broomsday_at_gmail.com> wrote:
> Jerome: You are certainly correct, I hadn't thought is was a single
> massively-dimensional bias. The memory thing makes much more sense now, as
> per your equation.
> Harish: Memory issues aside then, I don't think you would ever be able to
> reach equilibrium with such a massive coordinate space to bias across. If
> you look at Jerome's equation it not only relates to memory usage, but is
> also proportional to how many timesteps you would need to sample
> everything. The unfortunate truth is that given current technology, brute
> forcing a problem like this (trying to explore over all the conformations)
> just isn't possible. You need to do something that is less explicit in
> it's biasing, like replica exchange or some other form of accelerated
> sampling that doesn't require the full exploration of the coordinate space.
> On Tue, Jul 22, 2014 at 12:56 PM, Jérôme Hénin <jerome.henin_at_ibpc.fr>
>> Aron, when you ahd 9 colvars, maybe they were not assigned a single ABF
>> On 22 July 2014 18:30, Aron Broom <broomsday_at_gmail.com> wrote:
>>> I've used an upwards of 9 colvars at once with no issue, and with a
>>> larger system size than that. Can you post the *.in file or whatever file
>>> contains all your colvars definitions? I would suspect something add with
>>> your colvars. One thing to test might be, when you are at the largest size
>>> that has worked (I guess 4 colvars?) try just duplicating all those colvar
>>> blocks and see if it runs. Maybe something is wrong with the 5th colvar
>>> definition you are entering.
>>> On Tue, Jul 22, 2014 at 12:23 PM, Harish Vashisth <
>>> harish.vashisth_at_gmail.com> wrote:
>>>> Dear All:
>>>> We have been trying to make use of many CVs (> 20) using abf/colvar
>>>> options in NAMD2.9. These are all backbone dihedral CVs defined
>>>> individually in multiple blocks of phi/psi, etc. We are able to run ABF
>>>> jobs fine up to 5 CVs, but including a sixth one or more does not work. The
>>>> error reported in the log file right after initialization of colvars module
>>>> FATAL ERROR: Memory allocation failed on processor 0.
>>>> Looking through previous posts, someone seemed to suggest that it is
>>>> likely occurring due to large system size as NAMD keeps a copy of the
>>>> system on processor 0?
>>>> In our case, the solvated system size is ~40,000 atoms. The error
>>>> occurs using NAMD2.9 on local workstation, our local CRAYXE6m-200 as well
>>>> as on Stampede (XSEDE resource.) Is there an upper limit on how many CVs
>>>> colvar module can handle?
>>>> Any suggestions would be helpful. Thanks.
>>> Aron Broom M.Sc
>>> PhD Student
>>> Department of Chemistry
>>> University of Waterloo
> Aron Broom M.Sc
> PhD Student
> Department of Chemistry
> University of Waterloo
This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:22:38 CST