Date: Mon, 20 Jan 2020 15:01:36 -0300
From: Alvaro Herrera <alvherre@2ndquadrant.com>
To: Tomas Vondra <tomas.vondra@2ndquadrant.com>
Cc: "tsunakawa.takay@fujitsu.com" <tsunakawa.takay@fujitsu.com>,
	"pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>
Subject: Re: SLRU statistics
Message-ID: <20200120180136.GA28405@alvherre.pgsql>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20200120163754.mbbkbhzaummlo47c@development>
User-Agent: Mutt/1.10.1 (2018-07-13)
Precedence: bulk

On 2020-Jan-20, Tomas Vondra wrote:

> On Mon, Jan 20, 2020 at 01:04:33AM +0000, tsunakawa.takay@fujitsu.com wrote:
> > From: Tomas Vondra <tomas.vondra@2ndquadrant.com>
> > > One of the stats I occasionally wanted to know are stats for the SLRU
> > > stats (we have couple of those - clog, subtrans, ...). So here is a WIP
> > > version of a patch adding that.
> > 
> > How can users take advantage of this information?  I think we also need
> > the ability to set the size of SLRU buffers.  (I want to be freed from
> > the concern about the buffer shortage by setting the buffer size to its
> > maximum.  For example, CLOG would be only 1 GB.)
> 
> You're right the users can't really take advantage of this - my primary
> motivation was providing a feedback for devs, benchmarking etc. That
> might have been done with DEBUG messages or something, but this seems
> more convenient.

I think the stats are definitely needed if we keep the current code.
I've researched some specific problems in this code, such as the need
for more subtrans SLRU buffers; IIRC it was pretty painful to figure out
what the problem was without counters, and it'd have been trivial with
them.

> I think it's unclear how desirable / necessary it is to allow users to
> tweak those caches. I don't think we should have a GUC for everything,
> but maybe there's some sort of heuristics to determine the size. The
> assumption is we actually find practical workloads where the size of
> these SLRUs is a performance issue.

I expect we'll eventually realize the need for changes in this area.
Either configurability in the buffer pool sizes, or moving them to be
part of shared_buffers (IIRC Thomas Munro had a patch for this.)
Example: SLRUs like pg_commit and pg_subtrans have higher buffer
consumption as the range of open transactions increases; for many users
this is not a concern and they can live with the default values.

(I think when pg_commit (née pg_clog) buffers were increased, we should
have increased pg_subtrans buffers to match.)

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services