From owner-chemistry@ccl.net Mon Oct 1 03:53:01 2007 From: "andras.borosy(!)givaudan.com" To: CCL Subject: CCL: Free solubility prediction software Message-Id: <-35269-070928124420-12575-VhHFY1DZjbVXN/y0zzwM2A**server.ccl.net> X-Original-From: andras.borosy|-|givaudan.com Content-Type: multipart/alternative; boundary="=_alternative 0057EE64C1257364_=" Date: Fri, 28 Sep 2007 18:00:25 +0200 MIME-Version: 1.0 Sent to CCL by: andras.borosy===givaudan.com This is a multipart message in MIME format. --=_alternative 0057EE64C1257364_= Content-Type: text/plain; charset="ISO-8859-1" Content-Transfer-Encoding: quoted-printable Try SPARC! http://www.epa.gov/Athens/learn2model/part-two/onsite/sparcproperties.htm Regards, Dr. Andr=E1s P=E9ter Borosy Scientific Modelling Expert Fragrance Research Givaudan Schweiz AG - Ueberlandstrasse 138 - CH-8600 - D=FCbendorf -= =20 Switzerland T:+41-44-824 2164 - F:+41-44-8242926 - http://www.givaudan.com "Ivanciuc, Ovidiu I. oiivanci#%#utmb.edu" =20 Sent by: owner-chemistry##ccl.net 20/08/2007 22:07 Please respond to "CCL Subscribers" To "Borosy, Andras " cc Subject CCL: Free solubility prediction software Sent to CCL by: "Ivanciuc, Ovidiu I." [oiivanci[]utmb.edu] >>Is there any free software to calculate the solubility of small organic=20 molecules? >>Any online web server on this point is also appreciated. Go to the Virtual Computational Chemistry Laboratory,=20 http://www.vcclab.org/ and use ALOGPS, http://www.vcclab.org/lab/alogps/ ALOGPS computes the aqueous solubility from E-state topological indices and artificial neural networks. The dataset used to develop the solubility QSPR model is available from http://www.vcclab.org/lab/alogps/logs.txt and you may experiment with other machine learning models, such as support vector machines - for a list of SVM software see the=20 review "Applications of Support Vector Machines in Chemistry" http://www.ivanciuc.org/Files/Reprint/Ivanciuc=5FSVM=5FCCR=5F2007=5F23=5F29= 1.pdf and http://www.support-vector-machines.org/ Regards, Ovidiu -=3D This is automatically added to each message by the mailing script =3D-http://www.ccl.net/cgi-bin/ccl/send=5Fccl=5Fmessagehttp://www.ccl.net/cgi-bin/ccl/send=5Fccl=5Fmessage Subscribe/Unsubscribe:=20 http://www.ccl.net/chemistry/sub=5Funsub.shtmlJob: http://www.ccl.net/jobs=20http://www.ccl.net/spammers.txt--=_alternative 0057EE64C1257364_= Content-Type: text/html; charset="ISO-8859-1" Content-Transfer-Encoding: quoted-printable
Try SPARC!

 http://www.epa.gov/Athens/learn2model/part-two/onsite/sparcproperties.htm<= /font>


Regards,

Dr. Andr=E1s P=E9ter Borosy
Scientific Modelling Expert
Fragrance Research
Givaudan Schweiz AG  -  Ueberlandstrasse 138  -  CH-8600  -  D=FCbendorf  -  Switzerland
T:+41-44-824 2164  -  F:+41-44-8242926    -  http:= //www.givaudan.com



"Ivanciuc, Ovidiu I. oiivanci#%#utmb.edu" <owner-chemistry##ccl.net>
Sent by: owner-chemistry##ccl.net

20/08/2007 22:07
Please respond to
"CCL Subscribers" <chemistry##ccl.net>
To
"Borosy, Andras " <andras.borosy##givaudan.com>
cc
Subject
CCL: Free solubility prediction soft= ware






Sent to CCL by: "Ivanciuc, Ovidiu I." [oiivanci[]utmb.edu]


>>Is there any free software to calculate the solubility of small organic molecules?
>>Any online web server on this point is also appreciated.

Go to the Virtual Computational Chemistry Laboratory, http://www.vcclab.org= /
and use ALOGPS, http://www.vcclab.org/lab/alogps/

ALOGPS computes the aqueous solubility from E-state topological indices
and artificial neural networks.

The dataset used to develop the solubility QSPR model is available from
http://www.vcclab.org/lab/alogps/logs.txt
and you may experiment with other machine learning models,
such as support vector machines - for a list of SVM software see the review=
"Applications of Support Vector Machines in Chemistry"
http://www.ivanciuc.org/Files/Reprint/Ivanciuc=5FSVM=5FCCR=5F2007=5F23=5F29= 1.pdf

and http://www.support-vector-machines.org/

Regards,
Ovidiu



-=3D This is automatically added to each message by the mailing script =3D-=      http://www.ccl.net/cgi-bin/ccl/send=5Fccl=5Fmessage
     http://www.ccl.net/cgi-bin/ccl/send=5Fccl=5Fmessage
     http://www.ccl.net/chemistry/sub=5Funsub.shtml

Search Messages: http://www.ccl.net/htdig  (login: ccl, Password: search)
     http://www.ccl.net/spammers.txt
--=_alternative 0057EE64C1257364_=-- From owner-chemistry@ccl.net Mon Oct 1 07:12:00 2007 From: "Bruno Catalanotti brucatal]![unina.it" To: CCL Subject: CCL: Mopac on SGI irix 6.5 Message-Id: <-35270-071001070523-27605-V9W5HHUpjRc0RBUzd8pU/A=server.ccl.net> X-Original-From: "Bruno Catalanotti" Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="US-ASCII" Date: Mon, 1 Oct 2007 12:33:04 +0200 MIME-Version: 1.0 Sent to CCL by: "Bruno Catalanotti" [brucatal * unina.it] Hello all, Does somebody know which version of Mopac is currently available for (academic) SGI users? I am currently using the Mopac 6.0, implemented in a commercial software. Thank you very much Bruno Catalanotti From owner-chemistry@ccl.net Mon Oct 1 09:11:01 2007 From: "David Blackwell dtblackwell||gmail.com" To: CCL Subject: CCL:G: Gaussian 03 MP2 performance Message-Id: <-35271-071001085055-29197-fQHTY0kxPjhhzYV5YFtj+g{:}server.ccl.net> X-Original-From: "David Blackwell" Date: Mon, 1 Oct 2007 08:50:51 -0400 Sent to CCL by: "David Blackwell" [dtblackwell|a|gmail.com] Hi all, I've been running some single point MP2 calculations on medium-sized organic molecules (1012 basis functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately powerful workstation (Core 2 Quad processor, 4 GB RAM), and have been somewhat disappointed with the times taken for each job - each MP2 job takes > 3 days. Combined geometry optimization/frequency calculations on the same molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these timings seem about right for such a system? Is there any good way of optimising Gaussian's MP2 performance? The input files I've been using have been along the lines of: %Chk=whatever (where 'whatever' is the checkpoint file from the B3LYP optimization) [hash] MP2/cc-pVTZ Geom=AllCheck Cheers, David Blackwell From owner-chemistry@ccl.net Mon Oct 1 09:58:01 2007 From: "Herbert Fruchtl herbert.fruchtl a st-andrews.ac.uk" To: CCL Subject: CCL:G: Gaussian 03 MP2 performance Message-Id: <-35272-071001094958-17925-+KPFZiRA9fP0qSCOfS6bQw]_[server.ccl.net> X-Original-From: Herbert Fruchtl Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset=ISO-8859-1; format=flowed Date: Mon, 01 Oct 2007 14:49:17 +0100 MIME-Version: 1.0 Sent to CCL by: Herbert Fruchtl [herbert.fruchtl(-)st-andrews.ac.uk] A few things to look out for: - Are you using all the cores on your computer? Specify %NProcShared=4 on a quad processor. - Give it as much memory as possible. While SCF and DFT sometimes get slower when you give it more memory (which is weird...), MP2 should always benefit from more memory, because you get away with less integral passes in the semidirect calculation. For SMP calculations, you specify the total memory (minus what you need for static variables, executable, OS, etc.), not the amount per core. Give it 3.5GB or so. MP2 does get expensive as the number of basis functions increases (scaling as n**5). If it's only single-point calculations that you need, you may want to look at "local" or "RI" implementations (you find them in Q-Chem, NWChem, MOLPRO, ORCA, and probably several other programs nowadays; some of them even have gradients, which used to be the problem until lately). HTH, Herbert David Blackwell dtblackwell||gmail.com wrote: > Sent to CCL by: "David Blackwell" [dtblackwell|a|gmail.com] > Hi all, > I've been running some single point MP2 calculations on medium-sized organic molecules (1012 basis > functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately powerful workstation (Core 2 > Quad processor, 4 GB RAM), and have been somewhat disappointed with the times taken for each job > - each MP2 job takes > 3 days. Combined geometry optimization/frequency calculations on the same > molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these timings seem about right for > such a system? Is there any good way of optimising Gaussian's MP2 performance? > > The input files I've been using have been along the lines of: > > %Chk=whatever (where 'whatever' is the checkpoint file from the B3LYP optimization) > [hash] MP2/cc-pVTZ Geom=AllCheck > > Cheers, > David Blackwell> > > -- Herbert Fruchtl EaStCHEM Fellow School of Chemistry University of St Andrews From owner-chemistry@ccl.net Mon Oct 1 11:22:01 2007 From: "Bachrach, Steven STEVEN.BACHRACH^_^Trinity.edu" To: CCL Subject: CCL:G: Gaussian 03 MP2 performance Message-Id: <-35273-071001110914-1473-gVcmctjJCAqJ8yt20fhfmA|-|server.ccl.net> X-Original-From: "Bachrach, Steven" Content-class: urn:content-classes:message Content-Transfer-Encoding: quoted-printable Content-Type: text/plain; charset="us-ascii" Date: Mon, 1 Oct 2007 09:36:36 -0500 MIME-Version: 1.0 Sent to CCL by: "Bachrach, Steven" [STEVEN.BACHRACH-x-Trinity.edu] Yes, these performances are reasonable - I see the same type of behavior on my intel dual core and AMD dual core machines. The problem is really two-fold: the differences in the basis set and the difference in the QM method. The cc-pvtz is much bigger than the 6-31g(d) set, and also the Pople basis sets use the same exponent for the s and p functions so there is some real speed up there, that is not true for the cc basis sets MP2 simply scales much worse than DFT methods with size of the molecule - you can think of DFT computations as being some fractional increases over HF, but MP2 is order of magnitude slower than HF. Steve -- Steven Bachrach, Chair ph: (210)999-7379 Department of Chemistry fax: (210)999-7569 Trinity University =20 1 Trinity Place San Antonio, TX 78212 steven.bachrach]~[trinity.edu =20 -----Original Message----- > From: owner-chemistry]~[ccl.net [mailto:owner-chemistry]~[ccl.net]=20 Sent: Monday, October 01, 2007 7:51 AM To: Bachrach, Steven Subject: CCL:G: Gaussian 03 MP2 performance Sent to CCL by: "David Blackwell" [dtblackwell|a|gmail.com] Hi all, I've been running some single point MP2 calculations on medium-sized organic molecules (1012 basis=20 functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately powerful workstation (Core 2=20 Quad processor, 4 GB RAM), and have been somewhat disappointed with the times taken for each job=20 - each MP2 job takes > 3 days. Combined geometry optimization/frequency calculations on the same=20 molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these timings seem about right for=20 such a system? Is there any good way of optimising Gaussian's MP2 performance? The input files I've been using have been along the lines of: %Chk=3Dwhatever (where 'whatever' is the checkpoint file from the B3LYP optimization) [hash] MP2/cc-pVTZ Geom=3DAllCheck Cheers, David Blackwell -=3D This is automatically added to each message by the mailing script = =3D-http://www.ccl.net/cgi-bin/ccl/send_ccl_messageSubscribe/Unsubscribe:=20Job: http://www.ccl.net/jobs=20http://www.ccl.net/spammers.txt From owner-chemistry@ccl.net Mon Oct 1 12:00:01 2007 From: "John McKelvey jmmckel() gmail.com" To: CCL Subject: CCL:G: Gaussian 03 MP2 performance Message-Id: <-35274-071001113610-9264-9PGASXBTRy3W0blhPcqplA=-=server.ccl.net> X-Original-From: "John McKelvey" Content-Type: multipart/alternative; boundary="----=_Part_15049_28542695.1191252956958" Date: Mon, 1 Oct 2007 11:35:51 -0400 MIME-Version: 1.0 Sent to CCL by: "John McKelvey" [jmmckel~!~gmail.com] ------=_Part_15049_28542695.1191252956958 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Content-Disposition: inline I completely agree with Herbert about using "RI" with MP2.. It makes an _ENORMOUS_ difference in timing with no significant loss in accuracy. ORCA is free for academics and very easy to use. It runs on Windoze, and in parallel on Linux. Cheers! John McKelvey On 10/1/07, Herbert Fruchtl herbert.fruchtl a st-andrews.ac.uk < owner-chemistry ~ ccl.net> wrote: > > > Sent to CCL by: Herbert Fruchtl [herbert.fruchtl(-)st-andrews.ac.uk] > A few things to look out for: > - Are you using all the cores on your computer? Specify %NProcShared=4 > on a quad processor. > - Give it as much memory as possible. While SCF and DFT sometimes get > slower when you give it more memory (which is weird...), MP2 should > always benefit from more memory, because you get away with less integral > passes in the semidirect calculation. For SMP calculations, you specify > the total memory (minus what you need for static variables, executable, > OS, etc.), not the amount per core. Give it 3.5GB or so. > > MP2 does get expensive as the number of basis functions increases > (scaling as n**5). If it's only single-point calculations that you need, > you may want to look at "local" or "RI" implementations (you find them > in Q-Chem, NWChem, MOLPRO, ORCA, and probably several other programs > nowadays; some of them even have gradients, which used to be the problem > until lately). > > HTH, > > Herbert > > David Blackwell dtblackwell||gmail.com wrote: > > Sent to CCL by: "David Blackwell" [dtblackwell|a|gmail.com] > > Hi all, > > I've been running some single point MP2 calculations on medium-sized > organic molecules (1012 basis > > functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately > powerful workstation (Core 2 > > Quad processor, 4 GB RAM), and have been somewhat disappointed with the > times taken for each job > > - each MP2 job takes > 3 days. Combined geometry optimization/frequency > calculations on the same > > molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these > timings seem about right for > > such a system? Is there any good way of optimising Gaussian's MP2 > performance? > > > > The input files I've been using have been along the lines of: > > > > %Chk=whatever (where 'whatever' is the checkpoint file from the B3LYP > optimization) > > [hash] MP2/cc-pVTZ Geom=AllCheck > > > > Cheers, > > David Blackwell> > > > > > > -- > Herbert Fruchtl > EaStCHEM Fellow > School of Chemistry > University of St Andrews> > > > ------=_Part_15049_28542695.1191252956958 Content-Type: text/html; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Content-Disposition: inline I completely agree with Herbert about using "RI" with MP2.. It makes an _ENORMOUS_ difference in timing with no significant loss in accuracy.  ORCA is free for academics and very easy to use. It runs on Windoze, and in parallel on Linux.

Cheers!

John McKelvey

On 10/1/07, Herbert Fruchtl herbert.fruchtl a st-andrews.ac.uk < owner-chemistry ~ ccl.net> wrote:

Sent to CCL by: Herbert Fruchtl [herbert.fruchtl (-)st-andrews.ac.uk]
A few things to look out for:
- Are you using all the cores on your computer? Specify %NProcShared=4
on a quad processor.
- Give it as much memory as possible. While SCF and DFT sometimes get
slower when you give it more memory (which is weird...), MP2 should
always benefit from more memory, because you get away with less integral
passes in the semidirect calculation. For SMP calculations, you specify
the total memory (minus what you need for static variables, executable,
OS, etc.), not the amount per core. Give it 3.5GB or so.

MP2 does get expensive as the number of basis functions increases
(scaling as n**5). If it's only single-point calculations that you need,
you may want to look at "local" or "RI" implementations (you find them
in Q-Chem, NWChem, MOLPRO, ORCA, and probably several other programs
nowadays; some of them even have gradients, which used to be the problem
until lately).

HTH,

   Herbert

David Blackwell dtblackwell||gmail.com wrote:
> Sent to CCL by: "David  Blackwell" [dtblackwell|a|gmail.com]
> Hi all,
> I've been running some single point MP2 calculations on medium-sized organic molecules (1012 basis
> functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately powerful workstation (Core 2
> Quad processor, 4 GB RAM), and have been somewhat disappointed with the times taken for each job
> - each MP2 job takes > 3 days. Combined geometry optimization/frequency calculations on the same
> molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these timings seem about right for
> such a system? Is there any good way of optimising Gaussian's MP2 performance?
>
> The input files I've been using have been along the lines of:
>
> %Chk=whatever (where 'whatever' is the checkpoint file from the B3LYP optimization)
> [hash] MP2/cc-pVTZ Geom=AllCheck
>
> Cheers,
> David Blackwell>
>
>

--
Herbert Fruchtl
EaStCHEM Fellow
School of Chemistry
University of St Andrews



E-mail to subscribers: CHEMISTRY ~ ccl.net or use:
      http://www.ccl.net/cgi-bin/ccl/send_ccl_message

E-mail to administrators: CHEMISTRY-REQUEST ~ ccl.net or use
      http://www.ccl.net/cgi-bin/ccl/send_ccl_message

Subscribe/Unsubscribe:
      

Before posting, check wait time at: http://www.ccl.net

Job: http://www.ccl.net/jobs
Conferences: http://server.ccl.net/chemistry/announcements/conferences/

Search Messages: http://www.ccl.net/htdig   (login: ccl, Password: search)
      http://www.ccl.net/spammers.txt

RTFI: http://www.ccl.net/chemistry/aboutccl/instructions/




------=_Part_15049_28542695.1191252956958-- From owner-chemistry@ccl.net Mon Oct 1 12:32:01 2007 From: "Utpal Sarkar utpalchemiitkgp/a\gmail.com" To: CCL Subject: CCL: Help for MD simulation of metal cluster on substrate Message-Id: <-35275-071001093033-8460-OdixjIiI30XsIYGC4FyTvg__server.ccl.net> X-Original-From: "Utpal Sarkar" Content-Type: multipart/alternative; boundary="----=_Part_369_16495159.1191245404397" Date: Mon, 1 Oct 2007 15:30:04 +0200 MIME-Version: 1.0 Sent to CCL by: "Utpal Sarkar" [utpalchemiitkgp[-]gmail.com] ------=_Part_369_16495159.1191245404397 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Content-Disposition: inline Dear CCL Users, I want to use free molecular dynamics code to do the simulation of metal cluster (say Iron or Na) on substrate (say silicon). I want to know which molecular dynamics code can do it (specially to optimize the structures)? I also want to know is there any option in the code where I can implement my own potential energy function (or force field)? I will be highly obliged if you can give me some advise. Thanking you in advance With best regards Utpal Sarkar ------=_Part_369_16495159.1191245404397 Content-Type: text/html; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Content-Disposition: inline Dear CCL Users,
               I want to use free molecular dynamics code to do the simulation of metal cluster (say Iron
or Na) on substrate (say silicon). I want to know which molecular dynamics code can do it (specially to
optimize the structures)? I also want to know is there any option in the code where I can implement
my own potential energy function (or force field)?

I will be highly obliged if you can give me some advise.

Thanking you in advance
With best regards
Utpal Sarkar ------=_Part_369_16495159.1191245404397-- From owner-chemistry@ccl.net Mon Oct 1 13:25:00 2007 From: "Daniil Bratashov dn2010_-_gmail.com" To: CCL Subject: CCL: Mopac on SGI irix 6.5 Message-Id: <-35276-071001121821-12788-eqbexNY9/giggKLUfHlBDg-,-server.ccl.net> X-Original-From: Daniil Bratashov Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset=US-ASCII Date: Mon, 1 Oct 2007 19:12:55 +0400 Mime-Version: 1.0 Sent to CCL by: Daniil Bratashov [dn2010|gmail.com] On Mon, 1 Oct 2007 12:33:04 +0200 "Bruno Catalanotti brucatal]![unina.it" wrote: > Hello all, > Does somebody know which version of Mopac is currently available for > (academic) SGI users? I am currently using the Mopac 6.0, implemented > in a commercial software. > Thank you very much > Bruno Catalanotti Version 7.0 http://www.webmo.net/support/mopac_sgi.html Commercial mopac2002 was available some time ago: http://www.fqs.pl/chemistry/mopac_2002/new_version Also look at http://openmopac.net/ for more information about current development. WBR, Daniil Bratashov. From owner-chemistry@ccl.net Mon Oct 1 14:04:01 2007 From: "Jeff Nauss jnauss(-)accelrys.com" To: CCL Subject: CCL: Accelrys Customer Training for Oct and Nov 2007 Message-Id: <-35277-071001135157-8397-6osOIYks6qFdJ0ALuMpjXQ_._server.ccl.net> X-Original-From: Jeff Nauss Content-Type: text/plain; charset="US-ASCII" Date: Mon, 1 Oct 2007 10:51:45 -0700 MIME-Version: 1.0 Sent to CCL by: Jeff Nauss [jnauss() accelrys.com] Accelrys Inc. are holding the following training workshops during October and November 2007. These events are designed to help you get more value > from your Accelrys software, helping you to better accomplish your research goals. Costs for the on-line presentations, the price is $175 per three hour session for all customers. ON-LINE SESSIONS Catalyst to Discovery Studio Migration 10 Oct (8 am PST, 4pm BST) Insight II to Discovery Studio Migration 17 Oct (7am PST) Introduction to Materials Studio 23 Oct (8 am PST, 4pm BST) QSAR in Materials Studio 24 Oct (8 am PST, 4pm BST) Introduction to Discovery Studio 7 Nov (8 am PST, 4pm BST) Insight II to Discovery Studio Migration 9 Nov (8 am PST, 4pm BST) Catalyst to Discovery Studio Migration 28 Nov (8 am PST, 4pm BST) With the release of Discovery Studio 2.0, more workshops will be added in the near future. Please check back at http://www.accelrys.com/services/training/general/calendar.html for the latest offerings. Further details for all courses may be found on our website at http://www.accelrys.com/services/training/general/calendar.html and http://www.accelrys.com/services/training/index.html. Questions regarding scheduling and content should be directed to workshops===accelrys.com -- Jeffrey L. Nauss, Ph.D. Lead Training Scientist Accelrys 10188 Telesis Court, Suite 100 San Diego, CA 92121-4779 Phone: +1-858-799-5555 Fax: +1-858-799-5100 http://www.accelrys.com/services/training/ From owner-chemistry@ccl.net Mon Oct 1 15:04:00 2007 From: "Jerome Kieffer jerome.Kieffer*|*terre-adelie.org" To: CCL Subject: CCL: Mopac on SGI irix 6.5 Message-Id: <-35278-071001150036-18846-RhV1HDCWQaLw7vKHD6EGkw]-[server.ccl.net> X-Original-From: Jerome Kieffer Content-Transfer-Encoding: quoted-printable Content-Type: text/plain; charset=ISO-8859-15 Date: Mon, 1 Oct 2007 21:00:29 +0200 Mime-Version: 1.0 Sent to CCL by: Jerome Kieffer [jerome.Kieffer*o*terre-adelie.org] On Mon, 1 Oct 2007 12:33:04 +0200 "Bruno Catalanotti brucatal]![unina.it" wrote: >=20 > Sent to CCL by: "Bruno Catalanotti" [brucatal * unina.it] >=20 > Hello all, > Does somebody know which version of Mopac is currently available for > (academic) SGI users? I am currently using the Mopac 6.0, implemented in > a commercial software. I managed to compile MOPAC7 from the fortran sources available from sourceforge on SGI http://downloads.sourceforge.net/mopac7/mopac7-1.10.tgz MOPAC7 is public domain. --=20 J=E9r=F4me KIEFFER : http://www.terre-adelie.org =C0 v=E9lo, prendre une rue =E0 contre-sens est moins dangeureux que prendre un boulevard dans le sens l=E9gal. =C0 qui la faute ? From owner-chemistry@ccl.net Mon Oct 1 15:47:01 2007 From: "Daniel Jana dfjana{:}gmail.com" To: CCL Subject: CCL:G: Gaussian 03 MP2 performance Message-Id: <-35279-071001145810-18380-9/BO6zVaZTSJV6IXPsFC1g-$-server.ccl.net> X-Original-From: Daniel Jana Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset=ISO-8859-1; format=flowed Date: Mon, 01 Oct 2007 19:00:28 +0100 MIME-Version: 1.0 Sent to CCL by: Daniel Jana [dfjana|-|gmail.com] Hello! David Blackwell dtblackwell||gmail.com wrote: > Sent to CCL by: "David Blackwell" [dtblackwell|a|gmail.com] > Hi all, > I've been running some single point MP2 calculations on medium-sized organic molecules (1012 basis > functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately powerful workstation (Core 2 > Quad processor, 4 GB RAM), and have been somewhat disappointed with the times taken for each job > - each MP2 job takes > 3 days. Combined geometry optimization/frequency calculations on the same > molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these timings seem about right for > such a system? Is there any good way of optimising Gaussian's MP2 performance? > There is also another problem that hasn't been discussed here. It seems to me that Gaussian does not parallelize well. I have made some tests on a cluster (using Linda for the parallelization) and running a job on 4 processors take only 1/3 of the time. I did these tests for DFT methods, not sure if the same happens in MP2, nor were the tests very thorough. If this value is more or less correct you are losing one processor out of four which also helps explain your results. If so, then choosing an alternative method that does scale better, could (other than the options given before) further improve your results. Daniel Jana PS - As anyone observed similar (or totally different) results from my own? From owner-chemistry@ccl.net Mon Oct 1 16:43:01 2007 From: "Prof. E.-U. Wuerthwein wurthwe- -uni-muenster.de" To: CCL Subject: CCL: Gaussian 03 MP2 performance Message-Id: <-35280-071001115104-16075-/W6zokgh1+wZ8/FT8mSVJw()server.ccl.net> X-Original-From: "Prof. E.-U. Wuerthwein" Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed Date: Mon, 1 Oct 2007 17:20:32 +0200 (MES) MIME-Version: 1.0 Sent to CCL by: "Prof. E.-U. Wuerthwein" [wurthwe=uni-muenster.de] Have you tried MP2(FullDirect) in combination with as much memory as you have (e.g %mem=300000000)? This is usually very efficient, provided your molecule is not too large. Ernst From owner-chemistry@ccl.net Mon Oct 1 17:27:00 2007 From: "Jeff Hammond jeff.science * gmail.com" To: CCL Subject: CCL: Aces II now supported by EMSL's Basis Set Exchange Message-Id: <-35281-071001154206-8371-iBf92X4AFubLX7Dx+3n0BQ-*-server.ccl.net> X-Original-From: "Jeff Hammond" Content-Disposition: inline Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset=ISO-8859-1 Date: Mon, 1 Oct 2007 13:37:50 -0500 MIME-Version: 1.0 Sent to CCL by: "Jeff Hammond" [jeff.science/a\gmail.com] Dear Aces II Users, As of today, the EMSL Basis Set Exchange supports the Aces II format. Bert de Jong is responsibly for the upgrade; I am merely the messanger. Jeff From owner-chemistry@ccl.net Mon Oct 1 18:15:01 2007 From: "Jeff Hammond jeff.science * gmail.com" To: CCL Subject: CCL:G: Gaussian parallelism and parallelism in general Message-Id: <-35282-071001180726-5496-iy2iebYC1nhQ5dP7TPRPiQ%server.ccl.net> X-Original-From: "Jeff Hammond" Content-Disposition: inline Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset=ISO-8859-1 Date: Mon, 1 Oct 2007 17:07:14 -0500 MIME-Version: 1.0 Sent to CCL by: "Jeff Hammond" [jeff.science..gmail.com] A study of Gaussian 98's parallel performance is available here: http://www.hpc.uio.no/notur/gaussbench/gaussbench.php Some of the conclusions they reach (taken verbatim from that site) are: ========================================================== HF jobs - Jobs run well in parallel with 2-4 CPUs, in some cases parallel performance is decent even with 8 CPUs. We recommend using 1-4 CPUs. DFT jobs - Jobs run fine in parallel with 2-4 CPUs, so again the recommendation is to use 1-4 CPUs. MP2 jobs - Parallel performance is poor or non-existing. Run as serial jobs, i.e. on a single CPU. CC jobs - Parallel performance is poor for these methods as well. Should be run as serial jobs, but even single CPU jobs have relatively low efficiency. ========================================================== To be fair, the problem with Gaussian on parallel machines isn't necessarily that the methods aren't scalable, or that the people writing the code don't know what they're doing, but rather that the parallel model they chose to work with is Linda: http://en.wikipedia.org/wiki/Linda_(coordination_language). The Linda model is not a message-passing model like most modern parallel codes, and should not scale well on Beowulf-type machines since it assumes shared virtual memory (SVM). Since SVM is disk in the case of Gaussian, the problem with performance may have more to do with bandwidth limits in the data pipeline from disk to processor than inherent failures of the model itself. Gaussian probably scales reasonably well on SMP machines, such as the SGI Altix when running in-core, or machines with very fast disks (fast-SFS) when running on disk, but more tightly-coupled hardware is always more expensive than the commodity counterpart. At some point Gaussian won't scale well even on an SMP/fast-SFS machine, and this is a fundamental problem with most quantum chemical methods, particularly those which rely on canonical molecular orbitals, which preclude spatially-oriented data decomposition or sparse algorithms. Getting scalability beyond dozens of processors for quantum chemical codes is very difficult and has, to my knowledge, only been achieve by two kinds of developers: those using Global Arrays (NWChem, MolPro, UTChem) and those named Curt Janssen (MPQC). Please correct me if I have omitted anyone deserving. Jeff From owner-chemistry@ccl.net Mon Oct 1 20:00:00 2007 From: "Alex A. Granovsky gran:-:classic.chem.msu.su" To: CCL Subject: CCL:G: Gaussian 03 MP2 performance Message-Id: <-35283-071001135852-11914-NoVL6820ztFLVAa9994j6w.:.server.ccl.net> X-Original-From: "Alex A. Granovsky" Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="koi8-r" Date: Mon, 1 Oct 2007 22:16:39 +0400 MIME-Version: 1.0 Sent to CCL by: "Alex A. Granovsky" [gran a classic.chem.msu.su] Hi, just use PC GAMESS. Best regards, Alex Granovsky ----- Original Message ----- > From: "David Blackwell dtblackwell||gmail.com" To: "Granovsky, Alex, A. " Sent: Monday, October 01, 2007 4:50 PM Subject: CCL:G: Gaussian 03 MP2 performance > > Sent to CCL by: "David Blackwell" [dtblackwell|a|gmail.com] > Hi all, > I've been running some single point MP2 calculations on medium-sized organic molecules (1012 basis > functions using the cc-pVTZ basis set) in Gaussian 03 on a moderately powerful workstation (Core 2 > Quad processor, 4 GB RAM), and have been somewhat disappointed with the times taken for each job > - each MP2 job takes > 3 days. Combined geometry optimization/frequency calculations on the same > molecules at the B3LYP/6-31G(d) level take around 7 hours. Do these timings seem about right for > such a system? Is there any good way of optimising Gaussian's MP2 performance? > > The input files I've been using have been along the lines of: > > %Chk=whatever (where 'whatever' is the checkpoint file from the B3LYP optimization) > [hash] MP2/cc-pVTZ Geom=AllCheck > > Cheers, > David Blackwell> > > > From owner-chemistry@ccl.net Mon Oct 1 20:47:00 2007 From: "frisch[*]gaussian.com (Michael Frisch)" To: CCL Subject: CCL:G: Gaussian parallelism and parallelism in general Message-Id: <-35284-071001204234-12258-SyRiFbMDgTgNYq5qCWjzkw~!~server.ccl.net> X-Original-From: frisch*gaussian.com (Michael Frisch) Content-Disposition: inline Content-Type: text/plain; charset=us-ascii Date: Mon, 1 Oct 2007 19:41:54 -0400 Mime-Version: 1.0 Sent to CCL by: frisch..gaussian.com (Michael Frisch) On Mon, Oct 01, 2007 at 05:07:14PM -0500, Jeff Hammond jeff.science * gmail.com wrote: > > Sent to CCL by: "Jeff Hammond" [jeff.science..gmail.com] > A study of Gaussian 98's parallel performance is available here: > > http://www.hpc.uio.no/notur/gaussbench/gaussbench.php > Note that these results were for Gaussian 98; parallelism is considerably better for all these cases in Gaussian 03. Larger HF and DFT calculations use 8 processors quite effectively and show further speedups going from 8 to 16 processors on SMPs. > > To be fair, the problem with Gaussian on parallel machines isn't > necessarily that the methods aren't scalable, or that the people > writing the code don't know what they're doing, but rather that the > parallel model they chose to work with is Linda: > http://en.wikipedia.org/wiki/Linda_(coordination_language). > The above is nonsense; the communication time is not what limits parallelism in Gaussian; it is limited by Amdahl's law (the percentage of the calculation time which is parallel). Measurements of the peformance of Linda show that it is comparable to other methods of communication, although I repeat: this is not relevent to the performance of Gaussian on clusters. Gaussian does not use Linda at all for parallelism on SMPs; within one shared-memory system we use OpenMP (thread-based parallelism), which has zero communication overhead since all processors have access to the same memory system. > The Linda model is not a message-passing model like most modern > parallel codes, and should not scale well on Beowulf-type machines > since it assumes shared virtual memory (SVM). Nonsense. Data is communicated via IP when needed. Moving a matrix > from one worker to another is no different from any other communication protocol using the same interconnect. > Since SVM is disk in > the case of Gaussian, the problem with performance may have more to do > with bandwidth limits in the data pipeline from disk to processor than > inherent failures of the model itself. No data communicated between workers is written to disk. It is communicated from one worker to another via IP over the network (or via infiniband, or whatever the interconnect is). > Gaussian probably scales > reasonably well on SMP machines, such as the SGI Altix when running > in-core, or machines with very fast disks (fast-SFS) when running on > disk, but more tightly-coupled hardware is always more expensive than > the commodity counterpart. At some point Gaussian won't scale well > even on an SMP/fast-SFS machine, and this is a fundamental problem > with most quantum chemical methods, particularly those which rely on > canonical molecular orbitals, which preclude spatially-oriented data > decomposition or sparse algorithms. > More nonsense. The only "data communication" at all on an Altix is when one processor adds up results from several others, accessing the shared memory -- technically, if this is not local to the processor, it goes over the SGI routers, which have extremely high bandwidth, but this is all transparent to Gaussian or any other program running on the Altix. No disk I/O is involved. Disk I/O speed is irrelevent to G03 performance for jobs such as a DFT energy+gradient, which are compute bound. It is somewhat relevent for large DFT frequency calculations which do some I/O -- exactly the same amount of I/O whether running on 1 processor or 16. It is certainly relevent to coupled-cluster calculations, which do a lot of I/O, again independent of the the number of CPUs used. > Getting scalability beyond dozens of processors for quantum chemical > codes is very difficult and has, to my knowledge, only been achieve by > two kinds of developers: those using Global Arrays (NWChem, MolPro, > UTChem) and those named Curt Janssen (MPQC). Please correct me if I > have omitted anyone deserving. > Actually, I could make Gaussian show 99x speedup on 100 processors in 5 minutes. All I would have to do is add a few long loops which made the calculations take much longer and then parallelize those loops. Of course, this would get the answer more slowly than the current code, but it would show a great parallel speedup. This is a silly example, but it illustrates why the slower a code is on 1 processor, the easier it is to get good parallel speedups. Conversely, if I could make the 2e integral calculation twice as fast, then the parallel speedups would become worse, because it would reduce the time in the parallel part of the calculation. However, the program would run any given calculation faster than before on any number of processors despite the worse parallel speedup. So it is useful to understand the parallel speedups of a code one has already decided to use, but it is misleading to compare parallel speedups of different codes without looking at the actual time to solution for representative calculations (being careful to compare with similar accuracy settings, etc.) Mike Frisch