From owner-nwchem-users@emsl.pnl.gov Sun Apr 10 14:14:45 2005 Received: from odyssey.emsl.pnl.gov (localhost [127.0.0.1]) by odyssey.emsl.pnl.gov (8.12.10/8.12.10) with ESMTP id j3ALEjLN025287 for ; Sun, 10 Apr 2005 14:14:45 -0700 (PDT) Received: (from majordom@localhost) by odyssey.emsl.pnl.gov (8.12.10/8.12.10/Submit) id j3ALEjKA025286 for nwchem-users-outgoing; Sun, 10 Apr 2005 14:14:45 -0700 (PDT) Date: Sun, 10 Apr 2005 17:11:16 -0400 From: "William F. Polik" Subject: RE: ga_create error in frequency calculation In-reply-to: <200504102033.j3AKXrX7018687@zwart.qtp.ufl.edu> To: So Hirata Cc: nwchem-users@emsl.pnl.gov Message-id: <6.2.1.2.2.20050410170120.02607008@webmail.hope.edu> MIME-version: 1.0 X-Mailer: QUALCOMM Windows Eudora Version 6.2.1.2 Content-type: text/plain; charset=us-ascii; format=flowed X-WM-Posted-At: hope.edu; Sun, 10 Apr 05 17:18:25 -0400 References: <6.2.1.2.2.20050410134712.0265b940@webmail.hope.edu> <200504102033.j3AKXrX7018687@zwart.qtp.ufl.edu> Sender: owner-nwchem-users@emsl.pnl.gov Precedence: bulk So, Thank you very much. Your explanation about having to keep previous T amplitudes in memory for the DIIS algorithm makes good sense. I have a couple of follow-up questions regarding your suggestions: (1) I am running a sequential (one processor) version of nwchem. The documentation suggests that in this case, the replicated io scheme reduces to the Exclusive Access File (EAF) io scheme. Is it more efficient to simply choose "io eaf"? Also, how would this compare to "io fortran"? (2) The documentation suggests that one can turn off DIIS by specifying "diis 0". It seems that this would greatly reduce the memory requirements. However, would this approach slow down the convergence of the T amplitudes excessively? Thank you for your advice, Will At 04:33 PM 4/10/2005, So Hirata wrote: >Thank you for the output files. I think I know why your calculation ran out >of memory. It did go through CCSDT single point at equilibrium geometry, but >it crashed at 0th and 3rd iterations of CCSDT freq calculations for 400 and >800 mb runs, respectively. This is because the freq calculation needs to be >run at C1 symmetry (because finite displacements will be given to geometry) >and so it needs more memory than the single point at equilibrium which can >take advantage of C2v. 800 mb job did allocate more memory to TCE so two >iterations were performed. DIIS needs storage of previous T amplitudes and >the first 5 iterations will proportionally increase the memory usage. So >what I suggest as the solutions are as follows: > >(1) Use "io replicated". Default is "io ga". The former make the previous T >amplitudes stored in disk. > >(2) Increase further the memory size to 800 mb * 5/2 = 2 GB at least. And/or >decrease the DIIS order to 4 or 3 by "diis 3". ================================ Dr. William F. Polik Hofma Professor of Chemistry Department of Chemistry Hope College 35 East 12th Street Holland, MI 49422-9000 USA polik@hope.edu http://www.chem.hope.edu/~polik tel: (616) 395-7639 fax: (616) 395-7118 ================================