The ABINIT Users Mailing List ( CLOSED )

["H.L.Shih" <hlshih@gate.sinica.edu.tw>]

Dear Marc,

Thanks for the information. However, I used to VASP code before, and now 
I turn to Abinit for phonon study of metalic oxide. Judge from previous 
experience, the cost of calculation using PAW and US are quite the same 
when lunch as paralle. 

I adopted the stable version 4.3.3 on ibm, as you suggeted, maybe I should 
turn to 4.4.x for testing.

You might be true if including more valence electrons, however, I found in 
Fe PAW-GGA potential, only 6 electrons are treated as valence. However, in 
TM type, 8 electrons are considered. And I do agree with you that the 
radii used to constructed the pseudopotential should afect the 
convergence, but this is only true when pseudopotential itself containing 
more nodes which required higher Ecut to have same accuracy criterion. I 
understand  comparing two set should be meaningless, but I just wondering 
if performance could be better than the spec I got.

I'll look forward the next release of abinit, say 4.5. Thanks for the 
effort in fixing bugs and tuning the performance. :-)

BR,
J
On Thu, 16 Dec 2004 marc.torrent@cea.fr wrote:

> Dear Jason,
> 
> The answer is quite simple:
> This is not a normal behaviour of the PAW feature.
> 
> Let us remind you that:
> 1- In PAW, there is an atomic part to compute in addition to the "plane 
> waves" part (the one of the norm-conserving case).
> 2- But the size of PW basis to get a good convergency are smaller in PAW. 
> So you  MUST  decrease the needed CPU time.
> 3- You do not have to compare PAW and MT with the same cut-off.
> 
> 4- Which version of Abinit do you use ? (this is a compulsory information 
> to give when posting on the forum). Abinit PAW features are in v4.4.x in 
> "beta test" stage (and very slow in earlier versions). This "beta test" 
> characteristic will be suppressed in the future v4.5 version (end of 
> January).
> A lot of bugs remains in the <=v4.4.x versions.
> In addition this new version will be (quite) bug free, faster and have new 
> capabilities such as forces and stresses.
> 
> 5- If you downloaded the PAW atomic data on the web site, you must have 
> noticed that:
> - the PAW file includes more valence electrons (16) than the MT one (8) 
> (more cpu time needed).
> - the PAW data might have core radii differents from the MT ones (it has a 
> great influence on the computation speed).
> - the present PAW atomic data (on the web site) are only examples and not 
> optimized. In the next version (end of January) will plan to include new 
> PAW optimized data (we are now working on it). 
> 
> 
> Regards,
> 
> Marc Torrent and Francois Jollet
> CEA-Bruyeres-le-Chatel - France
> 
>  
> 
> 
> ----------------------------------------------------------
> # From: "H.L.Shih" <hlshih@gate.sinica.edu.tw>
> # To: forum@abinit.org
> # Subject: PAW pseudopotential
> # Date: Thu, 16 Dec 2004 21:20:51 +0800 (CST)
> 
> 
> Dear Abinit Users,
> 
> I tried to adopt PAW pseudopotential for simple Fe bulk, but found that
> the cost is about ten times slower comparing with ordinary TM type
> pseudopotential.
> 
> Table I. validation of bcc-Fe on ibm, pwr4 with two processors:
> =========================
> pot type cpu time
> (sec.)
> -------------------------
> TM 5.4
> PAW 50.7
> =========================
> 
> is there any parameter I can tune to make it faster? or this is quite
> normal when dealing with PAW. Thanks in advance.
> 
> BR,
> J
> 
> 

-- 
---------------------------------
Jason Shih
Academia Sinica Computer Center
No.128, Sec. 2, Academia Rd., 
Nangang District, Taipei City 115,Taiwan (R.O.C.)
Tel: +886-2-27899960
Fax: +886-2-27899949
---------------------------------