On Wed, 2006-11-22 at 11:02, Jeff Frost wrote:
> On Wed, 22 Nov 2006, Joshua D. Drake wrote:
>
> >> I could only find the 6 disk RAID5 numbers in the archives that were run with
> >> bonnie++1.03. Have you run the RAID10 tests since? Did you settle on 6 disk
> >> RAID5 or 2xRAID1 + 4XRAID10?
> >
> > Why not 6 drive raid 10? IIRC you need 4 to start RAID 10 but only pairs
> > after that.
>
> A valid question. Does the caching raid controller negate the desire to
> separate pg_xlog from PGDATA?
I remember seeing something on the list a while back that having
separate file systems was as important as having separate disks / arrays
for pg_xlog and PGDATA.
Something about the linux on the machine under test being better at
ordering of writes if they were to two separate file systems. Of
course, the weird thing is how counter intuitive that is, knowing that
the heads will have to move from one partition to another on a single
disk.
but on a multi-disk RAID10, it starts to make sense that the writes to
pg_xlog and the writes to data would likely be happening to different
drives at once, and so having them be on separate file systems would
make it faster if the kernel was better at handling the ordering that
way.
It's worth looking into at least.
Oh, and another vote of confidence for the LSI based controllers. I've
had good luck with both the "genuine" article from LSI and the Dell
aftermarket ones. Avoid the Dell - Adaptec controllers like the
plague. If you're lucky being slow is the only problem you'll have with
those.
I'm really hoping to spec out a data warehouse machine here in the next
year with lots of drives and an Areca or LSI controller in it... This
thread (and all the ones that have come before it) has been most useful
and will be archived.