Date: Mon, 26 Feb 2024 16:05:05 +0100
From: Christoph Berg <myon@debian.org>
To: Michael Paquier <michael@paquier.xyz>
Cc: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>,
	Alvaro Herrera <alvherre@alvh.no-ip.org>,
	Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>,
	vignesh C <vignesh21@gmail.com>,
	torikoshia <torikoshia@oss.nttdata.com>,
	PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Subject: Re: Printing backtrace of postgres processes
Message-ID: <Zdyooa_6QkCP3Lq7@msg.df7cb.de>
References: <ZcXeuOl9qy7YA9RH@paquier.xyz>
 <202402090848.j6evnpunbehe@alvherre.pgsql>
 <CAExHW5ugP_fAXtZhyES4zW5OtYn8J5e=eDZbSCCVM2ekASQRSg@mail.gmail.com>
 <Zca-C0rND4V0ZWI6@paquier.xyz>
 <Zdi8Qy-32rSrhy_S@msg.df7cb.de>
 <ZdksZzw6hrxFFXO0@paquier.xyz>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ZdksZzw6hrxFFXO0@paquier.xyz>
Archived-At: 
 <https://www.postgresql.org/message-id/Zdyooa_6QkCP3Lq7%40msg.df7cb.de>
Precedence: bulk

Re: Michael Paquier
> Something like this can be measured with a bunch of concurrent
> connections attempting connections and a very high rate, like pgbench
> with an empty script and -C, for local connections.

I tried that now. Mind that I'm not a benchmarking expert, and there's
been quite some jitter in the results, but I think there's a clear
trend.

Current head without and with the v28 patchset.
Command line:
pgbench -n -C -c 20 -j 20 -f empty.sql -T 30 --progress=2
empty.sql just contains a ";"
model name: 13th Gen Intel(R) Core(TM) i7-13700H

head:
tps = 2211.289863 (including reconnection times)
tps = 2113.907588 (including reconnection times)
tps = 2200.406877 (including reconnection times)
average: 2175

v28:
tps = 1873.472459 (including reconnection times)
tps = 2068.094383 (including reconnection times)
tps = 2196.890897 (including reconnection times)
average: 2046

2046 / 2175 = 0.941

Even if we regard the 1873 as an outlier, I've seen many vanilla runs
with 22xx tps, and not a single v28 run with 22xx tps. Other numbers I
collected suggested a cost of at least 3% for the feature.

Christoph