MIME-Version: 1.0
References: <20210719195950.gavgs6ujzmjfaiig@alap3.anarazel.de> <20230719085236.jltxal2eztqrprfg@awork3.anarazel.de>
In-Reply-To: <20230719085236.jltxal2eztqrprfg@awork3.anarazel.de>
From: John Naylor <john.naylor@enterprisedb.com>
Date: Tue, 8 Aug 2023 18:18:51 +0700
Message-ID: <CAFBsxsGdLkfqPdCFzQS-G7Gp6GXT44gT9M=3NYP67Dr3e_gWAA@mail.gmail.com>
Subject: Re: Avoid stack frame setup in performance critical routines using
 tail calls
To: Andres Freund <andres@anarazel.de>
Cc: David Rowley <dgrowleyml@gmail.com>, pgsql-hackers@postgresql.org, 
	Robert Haas <robertmhaas@gmail.com>, Michael Paquier <michael.paquier@gmail.com>, 
	Tomas Vondra <tv@fuzzy.cz>
Content-Type: multipart/alternative; boundary="000000000000bec4cb06026787b3"
Archived-At: <https://www.postgresql.org/message-id/CAFBsxsGdLkfqPdCFzQS-G7Gp6GXT44gT9M%3D3NYP67Dr3e_gWAA%40mail.gmail.com>
Precedence: bulk

--000000000000bec4cb06026787b3
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Wed, Jul 19, 2023 at 3:53=E2=80=AFPM Andres Freund <andres@anarazel.de> =
wrote:
>
> Hi,
>
> David and I were chatting about this patch, in the context of his bump
> allocator patch.  Attached is a rebased version that is also split up
into two
> steps, and a bit more polished.

Here is a quick test -- something similar was used to measure the slab
improvements last cycle. With radix tree v37 0001-0011 from [1],

create extension bench_radix_tree;
select avg(load_ms) from generate_series(1,100) x(x), lateral (select *
from bench_load_random_int(100 * 1000 * (1+x-x))) a;

The backend was pinned and turbo off. Perf runs were separate from timed
runs. I included 0002 for completeness.

v37
         avg
---------------------
 27.0400000000000000

  32.42%  postgres  bench_radix_tree.so  [.] rt_recursive_set
  21.60%  postgres  postgres             [.] SlabAlloc
  11.06%  postgres  [unknown]            [k] 0xffffffff930018f7
  10.49%  postgres  bench_radix_tree.so  [.] rt_extend_down
   7.07%  postgres  postgres             [.] MemoryContextAlloc
   4.83%  postgres  bench_radix_tree.so  [.] rt_node_insert_inner
   2.19%  postgres  bench_radix_tree.so  [.] rt_grow_node_48
   2.16%  postgres  bench_radix_tree.so  [.] rt_set.isra.0
   1.50%  postgres  bench_radix_tree.so  [.] MemoryContextAlloc@plt

v37 + palloc sibling calls
         avg
---------------------
 26.0700000000000000

v37 + palloc sibling calls + opt aset
         avg
---------------------
 26.0900000000000000

  33.78%  postgres  bench_radix_tree.so  [.] rt_recursive_set
  23.04%  postgres  postgres             [.] SlabAlloc
  11.43%  postgres  [unknown]            [k] 0xffffffff930018f7
  11.05%  postgres  bench_radix_tree.so  [.] rt_extend_down
   5.52%  postgres  bench_radix_tree.so  [.] rt_node_insert_inner
   2.47%  postgres  bench_radix_tree.so  [.] rt_set.isra.0
   2.30%  postgres  bench_radix_tree.so  [.] rt_grow_node_48
   1.88%  postgres  postgres             [.] MemoryContextAlloc
   1.44%  postgres  bench_radix_tree.so  [.] MemoryContextAlloc@plt

It's nice to see MemoryContextAlloc go down in the profile.

[1]
https://www.postgresql.org/message-id/CAD21AoA3gS45DFMOyTE-Wm4fu+BYzsYPVcSM=
YggLxwm40cGHZg@mail.gmail.com

--
John Naylor
EDB: http://www.enterprisedb.com

--000000000000bec4cb06026787b3
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><br>On Wed, Jul 19, 2023 at 3:53=E2=80=AFPM Andres Freund =
&lt;<a href=3D"mailto:andres@anarazel.de">andres@anarazel.de</a>&gt; wrote:=
<br>&gt;<br>&gt; Hi,<br>&gt;<br>&gt; David and I were chatting about this p=
atch, in the context of his bump<br>&gt; allocator patch.=C2=A0 Attached is=
 a rebased version that is also split up into two<br>&gt; steps, and a bit =
more polished.<br><br>Here is a quick test -- something similar was used to=
=C2=A0measure the slab improvements last cycle. With radix tree v37 0001-00=
11 from [1],<div><br></div><div>create extension bench_radix_tree;<br><div>=
select avg(load_ms) from generate_series(1,100) x(x), lateral (select * fro=
m bench_load_random_int(100 * 1000 * (1+x-x))) a;<br></div><div><br></div><=
div>The backend was pinned and turbo off. Perf runs were separate from time=
d runs. I included 0002 for completeness.</div><div><br></div><div>v37<br>=
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0avg =C2=A0 =C2=A0 =C2=A0 =C2=A0 <br>-----=
----------------<br>=C2=A027.0400000000000000<br><br>=C2=A0 32.42% =C2=A0po=
stgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_recursive_set<br>=C2=A0 21.60=
% =C2=A0postgres =C2=A0postgres =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 [=
.] SlabAlloc<br>=C2=A0 11.06% =C2=A0postgres =C2=A0[unknown] =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0[k] 0xffffffff930018f7<br>=C2=A0 10.49% =C2=A0po=
stgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_extend_down<br>=C2=A0 =C2=A07=
.07% =C2=A0postgres =C2=A0postgres =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=
=A0 [.] MemoryContextAlloc<br>=C2=A0 =C2=A04.83% =C2=A0postgres =C2=A0bench=
_radix_tree.so =C2=A0[.] rt_node_insert_inner<br>=C2=A0 =C2=A02.19% =C2=A0p=
ostgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_grow_node_48<br>=C2=A0 =C2=
=A02.16% =C2=A0postgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_set.isra.0<b=
r>=C2=A0 =C2=A01.50% =C2=A0postgres =C2=A0bench_radix_tree.so =C2=A0[.] Mem=
oryContextAlloc@plt<br><br>v37 + palloc sibling calls<br>=C2=A0 =C2=A0 =C2=
=A0 =C2=A0 =C2=A0avg =C2=A0 =C2=A0 =C2=A0 =C2=A0 <br>---------------------<=
br>=C2=A026.0700000000000000<br><br>v37 + palloc sibling calls + opt aset<b=
r>=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0avg =C2=A0 =C2=A0 =C2=A0 =C2=A0 <br>---=
------------------<br>=C2=A026.0900000000000000<br><br>=C2=A0 33.78% =C2=A0=
postgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_recursive_set<br>=C2=A0 23.=
04% =C2=A0postgres =C2=A0postgres =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0=
 [.] SlabAlloc<br>=C2=A0 11.43% =C2=A0postgres =C2=A0[unknown] =C2=A0 =C2=
=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0[k] 0xffffffff930018f7<br>=C2=A0 11.05% =C2=
=A0postgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_extend_down<br>=C2=A0 =
=C2=A05.52% =C2=A0postgres =C2=A0bench_radix_tree.so =C2=A0[.] rt_node_inse=
rt_inner<br>=C2=A0 =C2=A02.47% =C2=A0postgres =C2=A0bench_radix_tree.so =C2=
=A0[.] rt_set.isra.0<br>=C2=A0 =C2=A02.30% =C2=A0postgres =C2=A0bench_radix=
_tree.so =C2=A0[.] rt_grow_node_48<br>=C2=A0 =C2=A01.88% =C2=A0postgres =C2=
=A0postgres =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 [.] MemoryContextAllo=
c<br>=C2=A0 =C2=A01.44% =C2=A0postgres =C2=A0bench_radix_tree.so =C2=A0[.] =
MemoryContextAlloc@plt<br></div><div><br></div><div>It&#39;s nice to see Me=
moryContextAlloc go down in the profile.</div><div><br></div><div>[1]=C2=A0=
<a href=3D"https://www.postgresql.org/message-id/CAD21AoA3gS45DFMOyTE-Wm4fu=
+BYzsYPVcSMYggLxwm40cGHZg@mail.gmail.com">https://www.postgresql.org/messag=
e-id/CAD21AoA3gS45DFMOyTE-Wm4fu+BYzsYPVcSMYggLxwm40cGHZg@mail.gmail.com</a>=
<br><br>--<br>John Naylor<br>EDB: <a href=3D"http://www.enterprisedb.com">h=
ttp://www.enterprisedb.com</a></div></div></div>

--000000000000bec4cb06026787b3--