[CPMD-list] LOOP WAS VECTORIZED

[Axel Kohlmeyer]

On Fri, 4 Apr 2008, Nora Mohamed wrote:

NM> Dear Dr. Kohlmeyer,

dear nora,

[...]

NM> I was alarmed when you said that I have to compensate for the 
NM> dispersion interaction or I will get worse results than those from 
NM> classical MD, which brings me to you question about the choice of 
NM> the project; My supervisors published a paper about methane under 
NM> high pressures but using classical MD, my job is to get better 
NM> results especially for low temperatures using QM simulation. So it 

how high is high pressure supposed to be? if very high, you probably
have to start with learning how to generate pseudopotentials, since 
the normal ones are usually not "hard" enough. pseudization radii must
not overlap. 

NM> is vital to get better results or else it would be pointless to do a 
NM> QM simulation. I was warned several times that DFT doesn't represent 
NM> van der Waal interaction very well but I thought in comparison to 
NM> other interactions but still better than Classical MD not worse. 

nope. in DFT with typical functionals, e.g. BLYP, there is _no_ 
binding between methane molecules. please see the attached graph
taken from the CPMD-test distribution.

NM> doesn't CPMD offer the correction for van der Waal interaction 
NM> through VDW CORRECTION in &VDW section, is that what you mean by 
NM> compensating for the dispersion interaction?

that would be one of the options. there are a couple more 
methods, some implemented in CPMD, some not (yet). 

if you going very high pressure, that actually may not apply, 
since you'll be on the repulsive branch of the potential.
the problem of the empirical correction is that it is empirical,
i.e. if you fit a lj-potential, you probably get as good a 
result for _much_ lower cost.

since you mention low temperature, that will open a whole new
dimension of dificulty, because then quantum effects of the 
hydrogen atom core will matter and you'd have to run something
like path-integrals on top of everything. which will increase
the computational cost by about another two orders of magnitude.

NM> I actually tried to do a wavefunction optimization for 32 H2S 
NM> molecules with Cutoff 100, but after two days I stopped the run. I 

h2s is another problem. you have to have a proper pseudopotential
for sulfur and use it right. a friend of mine ran variable cell
CPMD calculations of h2s and told me some scary stories...

NM> thought that I must have done something wrong that made it take so 
NM> long, after all most people in this mailing list work on desktops 

you're mistaken there. most CPMD production runs are run on cluster
or larger parallel machines. in our group we rarely run on less than
32 cpus and frequently several hundreds up to occasionally a couple
of thousands when run on a bluegene/l machine. those people rarely
ask questions here (some of them should ;) ). most people posting 
questions here are beginners like you. 

[...]

NM> concerning grid computing, you said it is pointless, I don't 
NM> understand why , I mean unicore is a grid system that has CPMD on it 
NM> and people do testing jobs on it.

the very idea of grid computing is to make better use of resources
by having jobs migrated to the next matching available machine. that
works particularly well for standalone workstations or "compute farms".
if you have a large group of peopl who share resouces because they
don't need their machine in full all the time, but occasionally could
use a little extra. grid computing is the perfect solution to get
most out of your hardware.

to run "production" CPMD jobs you need a cluster or supercomputer 
and those are usually already (over-)loaded by regular users. a 
grid-interface just makes it more complicated to use, if you know 
how to automate your jobs and handle your job sumissions and data 
transfers efficiently, that is. unless we have so many similar 
enough clusters that it is efficient to migrate jobs around to the
next free slot on the next free cluster, grid for parallel jobs is
mostly hype and wishful thinking. every machine so far that i was
running on, needed some local tweaking and special scripts and
executables to get the most out of the available CPU budget.

in my opinion the largest advantage of, for example, teragrid in the 
USA is that we now have a coordinated management of computing resources 
throughout the network. that reduces the administrative overhead 
for the user enormously.

NM> I am sorry for the long email. Thank you Dr Kohlmeyer for your help.

well, mine wasn't exactly shorter. ;-).

cheers,
   axel.


NM> Best Regards,
NM> Nora M. Hassan
NM> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pot_ch4-ch4.jpg
Type: image/jpeg
Size: 52684 bytes
Desc: 
Url : http://cpmd.org/pipermail/cpmd-list/attachments/20080404/a9b82255/attachment-0001.jpg