Date: Tue, 28 Jan 1997 14:59:30 -0600
Sender: "SPSSX(r) Discussion" <SPSSX-L@UGA.CC.UGA.EDU>
From: Jonathan Fry <jon@SPSS.COM>
Organization: SPSS Inc.
Subject: Re: Memory limits of SPSS 6.1
Content-Type: text/plain; charset=us-ascii
Markus Quandt wrote:
> Hello everybody,
> I'm trying to run a complex macro which includes calculating maximum
> likelihood estimates and is supposed to operate on VERY large matrices
> (about 480 non-redundant parameters are to be estimated; the analysis
> will include about 75.000 valid cases).
> I'm aware that SPSS (6.1.3) is not the optimal platform to process this
> kind of problem (the matrix language operations are quite slow) but I am
> reluctant to switch to GAUSS or something alike because that would
> require conversion of the data and lots of time to find my way through a
> unknown system.
> Ok, this is what I get after setting both WORKSPACE and MXMEMORY to 32
> Number of locations requested = 41904889
> Number of locations available = 4178001
> This seems to tell me that the operation wants approx. 10 times the
> amount of memory that is available now, right? This would make 340
> megabytes or so.
> Now (finally) my question:
> Will a Win95-PC with enough free harddisk space (probably >> 340 MB
> required?!) but only 16MB of real memory allow SPSS to run the macro
> when WORKSPACE and MXMEMORY are set properly? That is, will virtual
> memory as provided by Win95 do?
> I can't simply try this myself because my machine has 'only' 200 MB
> harddisk space left. But getting at an appropriate machine could be
> possible, while 'outsourcing' to the Unix machines in the local
> computing center is a little involved and won't necessarily be faster
> (in terms of my own production cycles).
The message shown probably reflet a situation where MATRIX attempted to
allocate memory for a single matrix and failed. The cell count shown is
a little high for a 75,000x480 matrix, but could be right if you've
rounded the dimensions down.
So ten times the memory may not do the job. You can find out by seeing
if your current analysis will run with a tenth of your cases.