I wrote:
> Once this is in shape, I'll do some performance testing.
On second thought, there's no point in waiting, especially if a
regression points to a design flaw.
I compiled patched postgres with HEAP_FSM_CREATION_THRESHOLD set to
32, then ran the attached script which populates 100 tables with
varying numbers of blocks. I wanted a test that created pages eagerly
and wrote to disk as little as possible. Config was stock, except for
fsync = off. I took the average of 10 runs after removing the slowest
and fastest run:
# blocks master patch
4 36.4ms 33.9ms
8 50.6ms 48.9ms
12 58.6ms 66.3ms
16 65.5ms 81.4ms
It seems under these circumstances a threshold of up to 8 performs
comparably to the master branch, with small block numbers possibly
faster than with the FSM, provided they're in shared buffers already.
I didn't bother testing higher values because it's clear there's a
regression starting around 10 or so, beyond which it helps to have the
FSM.
A case could be made for setting the threshold to 4, since not having
3 blocks of FSM in shared buffers exactly makes up for the 3 other
blocks of heap that are checked when free space runs out.
I can run additional tests if there's interest.
-John Naylor
