From chemistry-request@server.ccl.net Sun Nov 3 08:55:12 2002 Received: from mail-b.bcc.ac.uk ([144.82.100.22]) by server.ccl.net (8.11.6/8.11.0) with ESMTP id gA3DtBw26540 for ; Sun, 3 Nov 2002 08:55:11 -0500 Received: from socrates-a.ucl.ac.uk by mail-b.bcc.ac.uk with SMTP (Mailer) with ESMTP; Sun, 3 Nov 2002 13:55:10 +0000 Received: (from uccatvm@localhost) by socrates-a.ucl.ac.uk (8.11.6+Sun/8.9.3) id gA3Dt9H12880; Sun, 3 Nov 2002 13:55:10 GMT From: uccatvm Message-Id: <200211031355.gA3Dt9H12880@socrates-a.ucl.ac.uk> Subject: Re:NWChem To: chemistry@ccl.net (CCL) Date: Sun, 3 Nov 2002 13:55:09 +0000 (GMT) Cc: T.vanMourik@ucl.ac.uk (Tanja van Mourik) X-Mailer: ELM [version 2.5 PL3] MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Hi Gregory and Dmitry, Gregory Shamov wrote: > > The main advantage of Gamess (Gaussian, Jaguar, Qchem, ...) over > > NWChem is that Gamess, in general, is stable, and NWChem, in general, > > is not. I don't have experience with Gamess, but I have used Gaussian and NWChem a lot. I have run NWChem on small and large machines, including Linux clusters, Suns, T3Es and SGIs, and I never experienced any stability problems. > > > > And there are some minor disadvantages of NWChem. First, it slow. Very > > slow. At single CPU it is about 2-4 times slower than Gamess. The > > NWChem team says it is because NWChem is designed to be very > > scalable code to use on hundreds and thousands CPUs. I believe it is > > true, but for commonly used small clusters (8-16 CPUs) Gamess > > outperforms NWChem. It is true that NWChem is slower than Gaussian on a single processor. However, the main advantage of NWChem is its, indeed, excellent scalability, and that holds for small as well as large clusters, for shared-memory as well as distributed-memory machines. In addition, because of its distributed data approach, it can handle larger calculations on parallel machines where programs that use a replicated-data approach are limited by the memory per processor (like Gaussian or Q-Chem). > > > > Moreover, NWChem (at least in some cases) running on the small cluster > > (say, 8 CPUs) gives no normal speedup painting! I.e., wall times for > > the same task are not smoothly decreasing with increasing CPU number, > > but are changing randomly. Perhaps in the case on 1000 CPUs cluster This is very strange, because I don't see that behaviour. I get very nice speedups with NWChem, on small as well as larger clusters. Running on 2 or 4 nodes, I generally get speedups of close to 2 and 4 on different architectures (that's for running ab initio - like SCF and MP2 - calculations). Even on Linux clusters connected by only fast ethernet I get speedups of 6.65 for running on 4 dual processors. With Gaussian, I can only run parallel within a dual processor box of my Linux cluster (because I don't have Linda), and I do not always get any speedup, as only part of Gaussian is parallelised. On the T3E, I could only use NWChem as I failed to run large (128-256-processor) calculations with Gaussian (because of the memory limitation). Thus, in general I use Gaussian for single-processor calculations, and NWChem for parallel calculations. > > +\- 8 CPUs means nothing, aha, aha, but for small cluster every CPU > > means a lot. :-) The same task calculated with Gamess on the same > > cluster scales properly. > > > > There are a lot of methods implemented in NWChem: DFT, CCSD, CPMD, > > classical MD, etc. But DFT (both grid-based and grid-free) and CCSD > > methods are implemented in the latest versions of Gamess too, it is > > possible to integrate Tinker with Gamess to do QM\MM and MD > > calculations, and there is standalone free and cool CPMD program. > > > > So i think you better use these codes unless you have 1000CPUs > > cluster. As I mentioned above, as I do not have any experience with Gamess, I cannot comment on its performance on small clusters, but in my experience NWChem does scale very nicely on these machines. In addition, the online manual (http://www.emsl.pnl.gov:2080/docs/nwchem/nwchem.html) is searchable, which I find very helpful, and although it doesn't have official support, the authors are very responsive to questions and problems on the nwchem mailing list. Thus, if you experience serious problems with NWChem, it may be worthwhile trying to get some feedback. Best wishes, Tanja -- ===================================================================== Tanja van Mourik Royal Society University Research Fellow Chemistry Department University College London phone: +44 (0)20-7679-4663 20 Gordon Street e-mail: work: T.vanMourik@ucl.ac.uk London WC1H 0AJ, UK home: tanja_van_mourik@btopenworld.com http://www.chem.ucl.ac.uk/people/vanmourik/index.html =====================================================================