Re[5]: LISTEN/NOTIFY ultra slow speed

Вторник, 15 марта 2022, 23:48 +03:00 от Арсен Арутюнян <arutar@bk.ru>:
 

I decided to rebuild pg 10 first.

I commented out the lines of code with duplicate checking

(/src/backend/commands/async.c)

 

//    if (AsyncExistsPendingNotify(channel, payload))

//        return;

 

and launched this version of pg into production.

This was the only way I could figure out if there was a significant difference in speed. And the results turned out to be very interesting, the query speed increased  by about two times. But unfortunately it's still far from the desired level

 

before the code changes

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 974,119 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 859,855 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 695,741 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 77,001 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 475,590 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 973,449 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 1039,825 мс (00:01,040)

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 1233,576 мс (00:01,234)

 

 

after the code changes

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 431,337 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 460,794 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 498,042 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 604,931 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 438,763 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 587,827 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 621,079 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 549,264 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 447,218 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 526,980 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48119697"}');

Время: 417,556 мс

 

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 185,374 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 554,529 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 654,769 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 577,187 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

Время: 246,831 мс

rmdatabase=# select pg_notify('rmdb_client_message','{"Command":239,"MsgId":"48117898"}');

 Время: 648,865 мс

 

tomorrow I will deal with the perf

 

--
Arsen Arutyunyan

 

 

Вторник, 15 марта 2022, 17:32 +03:00 от Tom Lane <tgl@sss.pgh.pa.us>:
 

Арсен Арутюнян <arutar@bk.ru> writes:
> i looked at the file "\src\backend\commands\async.c"
> there is a check for duplicate messages

Yup.

> Is this slowdown related to this check?

[ shrug... ] You've provided no evidence either for or against that
possibility.

It seems a little unlikely that you could have a lot of pending messages
without also seeing data accumulating in pg_notify/, but on the other hand
it's also hard to think of other places that time could be getting spent
if the queue is always empty. Can you use perf(1) or some such tool to
try to identify where the time is being spent? There's some advice
about that here:

https://wiki.postgresql.org/wiki/Profiling_with_perf

> Is it possible to skip the check if I'm absolutely sure the message is unique?

Not a good idea. You could upgrade to PG 13 or later, where that
code has been improved.

regards, tom lane