Re: PoC: Partial sort

From: Marti Raudsepp <marti(at)juffo(dot)org>
To: Alexander Korotkov <aekorotkov(at)gmail(dot)com>
Cc: Robert Haas <robertmhaas(at)gmail(dot)com>, Andreas Karlsson <andreas(at)proxel(dot)se>, David Rowley <dgrowleyml(at)gmail(dot)com>, pgsql-hackers <pgsql-hackers(at)postgresql(dot)org>, Andres Freund <andres(at)2ndquadrant(dot)com>, Martijn van Oosterhout <kleptog(at)svana(dot)org>
Subject: Re: PoC: Partial sort
Date: 2014-02-12 21:54:08
Message-ID: CABRT9RDLPxONHSR9BMuZpdkRf1Ww=RHvQdJ1_usGer9CO6yWGQ@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

On Mon, Feb 10, 2014 at 8:59 PM, Alexander Korotkov
<aekorotkov(at)gmail(dot)com> wrote:
> Done. Patch is splitted.

Thanks!

I think the 1st patch now has a bug in initial_cost_mergejoin; you
still pass the "presorted_keys" argument to cost_sort, making it
calculate a partial sort cost, but generated plans never use partial
sort. I think 0 should be passed instead. Patch attached, needs to be
applied on top of partial-sort-basic-1 and then reverse-applied on
partial-sort-merge-1.

With partial-sort-basic-1 and this fix on the same test suite, the
planner overhead is now a more manageable 0.5% to 1.3%; one test is
faster by 0.5%. Built with asserts disabled, ran on Intel i5-3570K. In
an effort to reduce variance, I locked the server and pgbench to a
single CPU core (taskset -c 3), but there are still noticeable
run-to-run differences, so these numbers are a bit fuzzy. The faster
result is definitely not a fluke, however; it happens every time.

> On Mon, Feb 10, 2014 at 2:33 PM, Marti Raudsepp <marti(at)juffo(dot)org> wrote:
>> AFAICT this only happens once per plan and the overhead is O(n) to the
>> number of pathkeys?

I was of course wrong about that, it also adds extra overhead when
iterating over the paths list.

----
Test "taskset -c 3 run2.sh" from
https://github.com/intgr/benchjunk/tree/master/partial_sort

Overhead percentages (between best of each 3 runs):
join5.sql 0.7
star5.sql 0.8
line5.sql 0.5
lim_join5.sql -0.5
lim_star5.sql 1.3
lim_line5.sql 0.5
limord_join5.sql 0.6
limord_star5.sql 0.5
limord_line5.sql 0.7

Raw results:
git 48870dd
join5.sql tps = 509.328070 (excluding connections establishing)
join5.sql tps = 509.772190 (excluding connections establishing)
join5.sql tps = 510.651517 (excluding connections establishing)
star5.sql tps = 499.208698 (excluding connections establishing)
star5.sql tps = 498.200314 (excluding connections establishing)
star5.sql tps = 496.269315 (excluding connections establishing)
line5.sql tps = 797.968831 (excluding connections establishing)
line5.sql tps = 797.011690 (excluding connections establishing)
line5.sql tps = 796.379258 (excluding connections establishing)
lim_join5.sql tps = 394.946024 (excluding connections establishing)
lim_join5.sql tps = 395.417689 (excluding connections establishing)
lim_join5.sql tps = 395.482958 (excluding connections establishing)
lim_star5.sql tps = 423.434393 (excluding connections establishing)
lim_star5.sql tps = 423.774305 (excluding connections establishing)
lim_star5.sql tps = 424.386099 (excluding connections establishing)
lim_line5.sql tps = 733.007330 (excluding connections establishing)
lim_line5.sql tps = 731.794731 (excluding connections establishing)
lim_line5.sql tps = 732.356280 (excluding connections establishing)
limord_join5.sql tps = 385.317921 (excluding connections establishing)
limord_join5.sql tps = 385.915870 (excluding connections establishing)
limord_join5.sql tps = 384.747848 (excluding connections establishing)
limord_star5.sql tps = 417.992615 (excluding connections establishing)
limord_star5.sql tps = 416.944685 (excluding connections establishing)
limord_star5.sql tps = 418.262647 (excluding connections establishing)
limord_line5.sql tps = 708.979203 (excluding connections establishing)
limord_line5.sql tps = 710.926866 (excluding connections establishing)
limord_line5.sql tps = 710.928907 (excluding connections establishing)

48870dd + partial-sort-basic-1.patch.gz + fix-cost_sort.patch
join5.sql tps = 505.488181 (excluding connections establishing)
join5.sql tps = 507.222759 (excluding connections establishing)
join5.sql tps = 506.549654 (excluding connections establishing)
star5.sql tps = 495.432915 (excluding connections establishing)
star5.sql tps = 494.906793 (excluding connections establishing)
star5.sql tps = 492.623808 (excluding connections establishing)
line5.sql tps = 789.315968 (excluding connections establishing)
line5.sql tps = 793.875456 (excluding connections establishing)
line5.sql tps = 790.545990 (excluding connections establishing)
lim_join5.sql tps = 396.956732 (excluding connections establishing)
lim_join5.sql tps = 397.515213 (excluding connections establishing)
lim_join5.sql tps = 397.578669 (excluding connections establishing)
lim_star5.sql tps = 417.459963 (excluding connections establishing)
lim_star5.sql tps = 418.024803 (excluding connections establishing)
lim_star5.sql tps = 418.830234 (excluding connections establishing)
lim_line5.sql tps = 729.186915 (excluding connections establishing)
lim_line5.sql tps = 726.288788 (excluding connections establishing)
lim_line5.sql tps = 728.123296 (excluding connections establishing)
limord_join5.sql tps = 383.484767 (excluding connections establishing)
limord_join5.sql tps = 383.021960 (excluding connections establishing)
limord_join5.sql tps = 383.722051 (excluding connections establishing)
limord_star5.sql tps = 414.138460 (excluding connections establishing)
limord_star5.sql tps = 414.063766 (excluding connections establishing)
limord_star5.sql tps = 416.130110 (excluding connections establishing)
limord_line5.sql tps = 706.002589 (excluding connections establishing)
limord_line5.sql tps = 705.632796 (excluding connections establishing)
limord_line5.sql tps = 704.991305 (excluding connections establishing)

Regards,
Marti

Attachment Content-Type Size
fix-cost_sort.patch text/x-patch 1.0 KB

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Bruce Momjian 2014-02-12 23:06:40 Re: Weaker shmem interlock w/o postmaster.pid
Previous Message Peter Eisentraut 2014-02-12 21:53:31 Re: Terminating pg_basebackup background streamer