On Tue, May 1, 2012 at 3:01 PM, Alvaro Herrera
<alvherre@commandprompt.com> wrote:
>> The system is supposed to be designed to avoid that; we try to hand out
>> pages with free space to different backends. One of the advantages of
>> the current page-at-a-time setup is that when there is *no* free space
>> according to the FSM, each such backend will create and fill its own
>> page independently. They do serialize on the extension lock, but the
>> intervening tuple additions are parallelized. We have to be careful to
>> not make that case worse in a multi-page-extension scheme.
>
> This didn't work all that well for toast tables that have lots of very
> large tuples inserted concurrently, at least with older releases. Each
> backend would acquire the extension lock many times per tuple inserted.
> The result was really high contention on the extension lock.
Hmm, this sounds like a good test case - a small number of SQL
statements triggering a large amount of relation extension work.
Using an unlogged table would probably make it easier to see the
relation-extension contention, too, since you'd get WALInsertLock
mostly out of the way.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company