On Tue, 9 Jan 2007, Jim C. Nasby wrote:
> On Thu, Dec 28, 2006 at 02:15:31PM -0800, Jeff Frost wrote:
>> When benchmarking various options for a new PG server at one of my clients,
>> I tried ext2 and ext3 (data=writeback) for the WAL and it appeared to be
>> fastest to have ext2 for the WAL. The winning time was 157m46.713s for
>> ext2, 159m47.098s for combined ext3 data/xlog and 158m25.822s for ext3
>> data=writeback. This was on an 8x150GB Raptor RAID10 on an Areca 1130 w/
>> 1GB BBU cache. This config benched out faster than a 6disk RAID10 + 2 disk
>> RAID1 for those of you who have been wondering if the BBU write back cache
>> mitigates the need for separate WAL (at least on this workload). Those are
>> the fastest times for each config, but ext2 WAL was always faster than the
>> other two options. I didn't test any other filesystems in this go around.
>
> Uh, if I'm reading this correctly, you're saying that WAL on a separate
> ext2 vs. one big ext3 with data=writeback saved ~39 seconds out of
> ~158.5 minutes, or 0.4%? Is that even above the noise for your
> measurements? I suspect the phase of the moon might play a bigger role
> ;P
That's what I thought too...cept I ran it 20 times and ext2 won by that margin
every time, so it was quite repeatable. :-/
--
Jeff Frost, Owner <jeff@frostconsultingllc.com>
Frost Consulting, LLC http://www.frostconsultingllc.com/
Phone: 650-780-7908 FAX: 650-649-1954