From gene@eastrg1.cray.com  Wed Apr 28 05:01:57 1993
Date: Wed, 28 Apr 93 09:01:57 EDT
From: gene@eastrg1.cray.com (Eugene Fleischmann)
Message-Id: <9304281301.AA06069@sodium.cray.com>
To: rbw@msc.edu
Subject: Re:  direct vs. in-core calculations


> From chemistry-request@ccl.net Mon Apr 26 19:10:44 1993
> To: DSMITH@uoft02.utoledo.edu, chemistry@ccl.net
> Subject: Re:  direct vs. in-core calculations
> Sender: chemistry-request@ccl.net
> 
> Doug,
> 
> If the memory is free and sufficient for the problem, you should
> always use INCORE (although I am not sure if this option is available
> on non-CRAY versions of the code). For closed shell systems memory
> requirements scale as N^4/4 where N is the number of basis functions.
> If you pay extra for memory then you need to consider the performance
> boost you get from running INCORE. On a CRAY-2 it can be as much as
> a factor of four. On the C90 it appears that it is a factor of three.
> At some point the DIRECT algorithm overcomes the INCORE algorithm,
> but this seems to be far beyond the basis function count that can 
> be fit INCORE on todays machines reasonably ... some 'memorable' data 
> points on the basis function memory relationship for closed shell
> systems are 100 basis function or fewer fit inside 100+ MWs (64 bit 
> words, 200 basis needs 200 MWs (64 bit words), 300 basis needs 1+ GWs 
> (again 64 bit words). Open shell systems need twice as much space.  
> 
> Sincerely,
> 
> Richard Walsh 
> Minnesota Supercomputer Center
> 
> 
> ---
> 

Richard,

I beg to differ with your numbers.  Using N^4/4 one calculates:

# functions 	memory
-----------	------
      100   ->   25 MW
      200   ->  400 MW
      300   -> 2025 MW

However, the Gaussian manual says, and my own experience support it,
that for closed shell RHF calculations memory scales as N^4/8 plus
some additional memory, say 500,000 words.  (G92 User's Guide p.88.)
My own experience has been that the additional .5 MW is too small,
a more realistic number is 100N*N.  Using this formula one calculates:

# functions 	memory
-----------	------
      100   ->   12.5 + 1 =   13.5 MW
      200   ->  200   + 4 =  204   MW
      300   -> 1012.5 + 9 = 1021.5 MW

A recent calculation of mine ran INCORE with 61 MW for 147 functions.
( 147  ->  58.4 + 2.2 = 60.6 MW )

Even this memory requirement for 300 functions (i.e. 2025 MW) is 
tractable on today's Cray systems.  I might also point out that use of 
the DIRECT keyword can still cause an INCORE calculation to run if 
sufficient memory has been provided using the %mem=???? directive in 
conjunction with the # QSUB lM ???mw option for NQS (if it is a batch 
job) and having no additional limits imposed on the job by the system.  
My opinion is that the INCORE is not warranted for production work at 
a site where no charge is incurred for memory, since the job may sit in 
a queue for some time, then fail due to lack of memory, whereas a DIRECT 
job will test for memory availability and run DIRECT if there is a lack 
of memory.  The trade off, of course, is that you will then pay for the 
EXTRA CPU time incurred running DIRECT instead of INCORE.

Regards,

Gene


++++++++++++++++++++++++++++++++++++++++
	Eugene D. Fleischmann, Ph.D.
	Computational Chemist
	Cray Research, Inc.
	(301)595-2695
	gene@eastrg2.cray.com
++++++++++++++++++++++++++++++++++++++++