On Tue, 08 Jun 2004 13:13:01 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>Possibly the relsize axis ought to be measured on a log scale, or
>something like that, but that didn't seem to work nicely when relsize
>approaches zero.
In my experiments I used log(relsize) on the x axis, and I don't think
that the graph looks unpleasant for small relsize. My thought was (and
is) that we are much more interested in whether relpages is 1/100, 1/10,
1, 10, 100 times effective_cache_size than whether it is relpages +/-
1000, 2000, 3000, ...
I played around with some numbers that could be considered fairly
realistic. You might want to look at the graphs I linked to in the
previous message or download http://www.pivot.at/pg/costsize.sxc.
But I think we are wasting too much effort. The graphs don't look too
different, whether you use relsize or relsize^2. Maybe relsize^3 is
optimal? Nobody knows. The important part of the patch is that the
result is scaled and shifted into the range 1 to random_page_cost.
Whatever you decide to do is ok with me.
Servus
Manfred