Re: auto-sizing wal_buffers

On Sat, Jan 15, 2011 at 2:34 PM, Josh Berkus <josh(at)agliodbs(dot)com> wrote:
> On 1/14/11 10:51 PM, Greg Smith wrote:
>>
>> !         Since the data is written out to disk at every transaction
>> commit,
>> !         the setting many only need to be be large enough to hold the
>> amount
>> !         of WAL data generated by one typical transaction.  Larger values,
>> !         typically at least a few megabytes, can improve write performance
>> !         on a busy server where many clients are committing at once.
>> !         Extremely large settings are unlikely to provide additional
>> benefit.
>
> I think we can be more specific on that last sentence; is there even any
> *theoretical* benefit to settings above 16MB, the size of a WAL segment?

I would turn it around and ask if there is any theoretical reason it
would not benefit?
(And if so, can they be cured soon?)

> Certainly there have been no test results to show any.

Did the tests show steady improvement up to 16MB and then suddenly
hit a wall? (And in which case, were they recompiled at a larger segment
size and repeated?) Or did improvement just peter out because 16MB is really
quite a bit and there was just no need for it to be larger independent
of segment size?

Cheers,

Jeff