I wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes:
>> One idea might be to run a whole bunch of queries and record all of
>> the planning times, and then run them all again and compare somehow.
>> Maybe the regression tests, for example.
> That sounds like something we could do pretty easily, though interpreting
> the results might be nontrivial.
I spent some time on this project. I modified the code to log the runtime
of standard_planner along with decompiled text of the passed-in query
tree. I then ran the regression tests ten times with cassert-off builds
of both current HEAD and HEAD+pathification patch, and grouped all the
numbers for log entries with identical texts. (FYI, there are around
10000 distinguishable queries in the current tests, most planned only
once or twice, but some as many as 2900 times.) I had intended to look
at the averages within each group, but that was awfully noisy; I ended up
looking at the minimum times, after discarding a few groups with
particularly awful standard deviations. I theorize that a substantial
part of the variation in the runtime depends on whether catalog entries
consulted by the planner have been sucked into syscache or not, and thus
that using the minimum is a reasonable way to eliminate cache-loading
effects, which surely ought not be considered in this comparison.
Here is a scatter plot, on log axes, of planning times in milliseconds
with HEAD (x axis) vs those with patch (y axis):
