Re: Terrible Write Performance of a Stored Procedure

Поиск
Список
Период
Сортировка
От Greg Smith
Тема Re: Terrible Write Performance of a Stored Procedure
Дата
Msg-id alpine.GSO.2.01.0906261816530.22738@westnet.com
обсуждение исходный текст
Ответ на Re: Terrible Write Performance of a Stored Procedure  (Scott Mead <scott.lists@enterprisedb.com>)
Ответы Re: Terrible Write Performance of a Stored Procedure
Список pgsql-performance
On Fri, 26 Jun 2009, Scott Mead wrote:

>     Having those settings enabled basically does the following:
>    " Do not complete the I/O for a commit until you have either commit_siblings commits also ready, or you have
waited.55 seconds." 
>
>   Basically, if you make 1 commit, you will sit there waiting until either 99 other commits take place, or ~ 1/2
secondgoes by. 

You're right that it should be removed, but this explanation is wrong.
The behavior as configured is actually "if there are >=100 other
transactions in progress, wait 0.1 second before committing after the
first one gets committed", in hopes that one of the other 100 might also
join along in the disk write.

Since in this case max_connections it set to 100, it's actually impossible
for the commit_delay/commit_siblings behavior to trigger give this
configuration.  That's one reason it should be removed.  The other is that
i general, if you don't exactly what you're doing, you shouldn't be
touching either parameters; they don't do what people expect them to and
it's extremely unlikely you'll encounter any of the rare use cases where
they might help.

I don't think any of the sync or write parameters have anything to do with
this problem though, it seems like a problem with the referential bits
taking too long to execute.

--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD

В списке pgsql-performance по дате отправления:

Предыдущее
От: Tom Lane
Дата:
Сообщение: Re: Nested Loop "Killer" on 8.1
Следующее
От: Tom Lane
Дата:
Сообщение: Re: Insert performance and multi-column index order