On Thu, 13 Mar 2008, Joshua D. Drake wrote:
> Greg Smith <gsmith@gregsmith.com> wrote:
>>>> wal_sync_method = open_sync
>>
>> There was a bug report I haven't had a chance to investigate yet that
>> suggested some recent Linux versions have issues when using
>> open_sync. I'd suggest popping that back to the default for now
>> unless you have time to really do a long certification process that
>> your system runs reliably with it turned on.
>
> Well the default would be ugly, that's fsync, fdatasync is probably a
> better choice in that case.
I haven't found fdatasync to be significantly better in my tests on Linux
but I never went out of my way to try and quantify it. My understanding
is that some of the write barrier implementation details on ext3
filesystems make any sync call a relatively heavy operation but I haven't
poked at the code yet to figure out why.
There are really some substantial gains for WAL-heavy loads under Linux
just waiting for someone to dig into this more. For example, I have a
little plan sitting here to allow opening the WAL files with noatime even
if the rest of the filesystem can't be mounted that way, which would
collapse one of the big reasons to use a separate WAL disk.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD