On Fri, 3 Mar 2023 at 15:17, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> (Is it worth carrying both "allocated words" and "nonzero words"
> fields to avoid useless memory-management effort? Dunno.)
It would have been a more appealing thing to do before Bitmapset
became a node type as we'd have had empty space in the struct to have
another 32-bit word on 64-bit builds.
> Another point here is that I'm pretty sure that just about all
> bitmapsets we deal with are only one or two words, so I'm not
> convinced you're going to get any performance win to justify
> the added management overhead.
It's true that the majority of Bitmapsets are going to be just 1 word,
but it's important to acknowledge that we do suffer in some more
extreme cases when Bitmapsets become large. Partition with large
numbers of partitions is one such case.
create table lp(a int) partition by list(a);
select 'create table lp'||x||' partition of lp for values
in('||x||');' from generate_series(0,9999)x;
\gexec
# cat bench.sql
select * from lp where a > 1 and a < 3;
$ pgbench -n -T 15 -f bench.sql postgres | grep tps
master:
tps = 28055.619289 (without initial connection time)
tps = 27819.235083 (without initial connection time)
tps = 28486.099808 (without initial connection time)
master + bms_no_trailing_zero_words.patch:
tps = 30840.840266 (without initial connection time)
tps = 29491.519705 (without initial connection time)
tps = 29471.083938 (without initial connection time)
(~6.45% faster)
Of course, it's an extreme case, I'm merely trying to show that
trimming the Bitmapsets down can have an impact in some cases.
David
