Mime-Version: 1.0
Content-Transfer-Encoding: quoted-printable
Content-Type: text/plain; charset=UTF-8; format=Flowed
Date: Fri, 27 Dec 2024 15:15:40 +0100
Message-Id: <D6MJOHS7HZ80.3B17NDGUV6T47@jeltef.nl>
Cc: "Pgsql Hackers" <pgsql-hackers@postgresql.org>
Subject: Re: RFC: Allow EXPLAIN to Output Page Fault Information
To: "Tom Lane" <tgl@sss.pgh.pa.us>, "torikoshia"
 <torikoshia@oss.nttdata.com>
From: "Jelte Fennema-Nio" <postgres@jeltef.nl>
References: <c20f6340eb26f3b736abc59471bfada8@oss.nttdata.com>
 <1614321.1735055528@sss.pgh.pa.us>
In-Reply-To: <1614321.1735055528@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/D6MJOHS7HZ80.3B17NDGUV6T47%40jeltef.nl>
Precedence: bulk

On Tue Dec 24, 2024 at 4:52 PM CET, Tom Lane wrote:
> torikoshia <torikoshia@oss.nttdata.com> writes:
>> I have attached a PoC patch that modifies EXPLAIN to include page fault=
=20
>> information during both the planning and execution phases of a query.=20
>
> Surely these numbers would be too unstable to be worth anything.

What makes you think that? I'd expect them to be similarly stable to the
numbers we get for BUFFERS. i.e. Sure they won't be completely stable,
but I expect them to be quite helpful when debugging perf issues,
because large numbers indicate that the query is disk-bound and small
numbers indicate that it is not.

These numbers seem especially useful for setups where shared_buffers is
significantly smaller than the total memory available to the system. In
those cases the output from BUFFERS might give the impression that that
you're disk-bound, but if your working set still fits into OS cache then
the number of page faults is likely still low. Thus telling you that the
numbers that you get back from BUFFERS are not as big of a problem as
they might seem.