MIME-Version: 1.0
References: 
 <OS9PR01MB121499E64344675A86179219CF579A@OS9PR01MB12149.jpnprd01.prod.outlook.com>
 <PH0PR08MB7020F6EA8877834CCDD5FABB8879A@PH0PR08MB7020.namprd08.prod.outlook.com>
 <408cd5617d954ba1aca99f26b4693fa2@localhost.localdomain>
 <SA1PR08MB702308B377B110460FCDADA18847A@SA1PR08MB7023.namprd08.prod.outlook.com>
 <4a3bfa370a7d48d1965b12572a0de0e5@localhost.localdomain>
 <PH0PR08MB7020956FB86755AD6B4E2DCF8844A@PH0PR08MB7020.namprd08.prod.outlook.com>
In-Reply-To: 
 <PH0PR08MB7020956FB86755AD6B4E2DCF8844A@PH0PR08MB7020.namprd08.prod.outlook.com>
From: Fujii Masao <masao.fujii@gmail.com>
Date: Fri, 20 Mar 2026 00:29:46 +0900
Message-ID: 
 <CAHGQGwGKAr7GFD5mS3aGi4W4KpmP9wVkbYRWEHOQM8u18qX9Cg@mail.gmail.com>
Subject: Re: Idea to enhance pgbench by more modes to generate data
 (multi-TXNs, UNNEST, COPY BINARY)
To: Boris Mironov <boris_mironov@outlook.com>
Cc: Madyshev Egor <E.Madyshev@ftdata.ru>,
	"Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>,
	"pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: 
 <https://www.postgresql.org/message-id/CAHGQGwGKAr7GFD5mS3aGi4W4KpmP9wVkbYRWEHOQM8u18qX9Cg%40mail.gmail.com>
Precedence: bulk

On Thu, Mar 12, 2026 at 11:28=E2=80=AFPM Boris Mironov
<boris_mironov@outlook.com> wrote:
>
> Hi Egor,
>
> Thank you very much for your time reviewing this patch
> and guiding through some of its inefficiencies.

I like the idea of improving the performance of the initial data load
in pgbench. That's definitely useful.

I ran some tests on my MacBook (server and pgbench on the same machine)
using different data loading options. Here are the results:

------------------------------------------------
[ Client / Text mode ]
pgbench -i -Idtg -s100
done in 13.38 s (drop tables 0.00 s, create tables 0.01 s, client-side
generate 13.38 s).

pgbench -i -Idtg -s1000
done in 151.81 s (drop tables 0.00 s, create tables 0.01 s,
client-side generate 151.81 s).

[ Client / Binary mode ]
pgbench -i -Idtc -s100
done in 18.32 s (drop tables 0.00 s, create tables 0.01 s, client-side
generate 18.31 s).

pgbench -i -Idtc -s1000
done in 204.48 s (drop tables 0.00 s, create tables 0.01 s,
client-side generate 204.47 s).

[ Server / generate_series ]
pgbench -i -IdtG -s100
done in 21.30 s (drop tables 0.00 s, create tables 0.00 s, server-side
generate 21.30 s).

pgbench -i -IdtG -s1000
done in 230.94 s (drop tables 0.00 s, create tables 0.01 s,
server-side generate 230.93 s).

[ Server / Unnest ]
pgbench -i -IdtU -s100
done in 23.16 s (drop tables 0.00 s, create tables 0.00 s, server-side
generate 23.16 s).

pgbench -i -IdtU -s1000
done in 249.08 s (drop tables 0.00 s, create tables 0.01 s,
server-side generate 249.07 s).
------------------------------------------------

In my tests, text mode was faster than binary mode. Also, on the server sid=
e,
generate_series() was faster than unnest().

So I'm wondering if there are specific conditions where binary mode or unne=
st()
performs better. If not, it may not be worth supporting these additional da=
ta
loading options...

Regards,


--=20
Fujii Masao