On Thursday, May 31, 2012 01:33:33 AM Fujii Masao wrote:
> On Wed, May 30, 2012 at 9:46 PM, Andres Freund <andres@2ndquadrant.com>
wrote:
> > On Tuesday, May 29, 2012 08:42:43 PM Andres Freund wrote:
> >> Patch attached.
> >
> > Imo this patch should be backported to 9.1, 9.0 doesn't use latches and
> > does not do explicit wakeup of the sender so its not applicable there.
> >
> > I can prepare a patch for 9.1 if people agree, there has been some amount
> > of change that won't make it apply cleanly.
>
> The patch wakes up walsender more frequently than now. Which leads
> walsender to send smaller WAL data packet more frequently, and furthermore
> which leads walreceiver to issue fsync more frequently. So I'm afraid that
> the patch makes the disk more busy and slows down the read-only query
> in the standby. I'm also afraid that frequent fsync calls degrade the
> performance
> of sync replication. So it's better to do benchmark to address the
> concerns.
I couldn't measure any significant difference in #fsyncs or replication speed
in a busy pgbench workload. If anything there were less, but the difference
was small. Both with synchronous_commit=on/off. I did *not* test sync repl.
Why would you expect a change? Walsender is only signalled if XLogWrite
actually flushed data to disk. Thats the point of the exercise. Thats normally
only done if the wal buffers are full or a commit record (+some other stuff)
requires the xlog to be flushed up to some point.
Greetings,
Andres
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services