On Sat, 16 Nov 2002, Stephan Szabo wrote:
> On Sun, 17 Nov 2002, pginfo wrote:
>
> > Hi,
> >
> > Stephan Szabo wrote:
> >
> > > On Sat, 16 Nov 2002, pginfo wrote:
> > >
> > > > Hi,
> > > >
> > > > Tom Lane wrote:
> > > >
> > > > > Josh Berkus <josh@agliodbs.com> writes:
> > > > > > Here's a question: is the total size of the column a good indicator of the
> > > > > > sort_mem required? Or does the rowsize affect it somehow?
> > > > >
> > > > > It will include all the data that's supposed to be output by the sort...
> > > > > both the key column(s) and the others.
> > > > >
> > > >
> > > > Hmm it is not clear for me.Let we have all data.
> > > > If I make sort by S.OP ( it is INT) it take < 6 sek for sort.
> > > > I think we move all this data anly the number of comparation is by INT. I think
> > > > the number of comparation
> > > > is ~ n * ln(n).
> > > > If we sort by S.IDS_xxx we have also n*ln(n) comparations but in
> > > > varchar(string).
> > > > I don't think that it can take 50 sek.
> > > >
> > > > Is it not so?
> > >
> > > Have you tried setting up another database in "C" locale and compared the
> > > timings there? I'd wonder if maybe there's some extra copying going on
> > > given the comments in varstr_cmp.
> >
> > No, I do not have any info about it.I will see if it is possible ( the data are not
> > so simple).
> > If it is possible I will make the tests.
> > Have no one that have 700K row in thow tables?
> > It is simple to test:
> > 1. Run query that returns ~700K rows from this tables.
> > 2. Make sort.
> >
> > It is interest only the sort time!
>
> I can make a table of 700k rows and test it (and am generating 700k of
> random varchar rows), but I wouldn't hold great hope that this is
> necessarily a valid test since possibly any of OS, configuration settings
> and actual data (width and values) might have an effect on the results.
On my not terribly powerful or memory filled box, I got a time of about
16s after going through a couple iterations of raising sort_mem and
watching if it made temp files (which is probably a good idea to check as
well). The data size ended up being in the vicinity of 100 meg in my
case.