Hi Jerry,
Jeremy is right about placing data on a separate array for increased
performance. If you want the best, then I'd place it on an NVRAM device. I
did say that the performance is negligible. This doesn't mean that there
isn't any benefit. It just means I wouldn't pay the extra price for what
little is gained.
I've attached a single page excerpt from some of the testing I've done in
the past. The server is a dual 2.0GHz Xeon. The client is a single 2.4GHz
Xeon. The OS was RedHat 9.0 with a new 2.4.20 kernel and XFS.
In this specific case, I used md RAID on 11 10K RPM Seagate FC drives. The
lower performance line created a single array and used an internal log. The
top line took the same 11 drives and created one 256MB RAID for metadata and
the rest of the drives was arrayed for data. Both RAIDs used RAID-5 on the
same physical, shared disks.
The point of all this was to address the initial question of whether or not
to use an external log on another partition. I'm not trying to tell you how
to design the fastest possible system. Economics are sometimes a factor.
However, you can expect some very definite performance gains over your
default config. Your call.
-Mike
-----Original Message-----
From: linux-xfs-bounce@xxxxxxxxxxx [mailto:linux-xfs-bounce@xxxxxxxxxxx] On
Behalf Of Jerry Haltom
Sent: Friday, January 02, 2004 8:26 AM
To: Jeremy Jackson
Cc: Wvoice; 'linux-xfs list'
Subject: Re: XFS external log questions
And here is exactly why I asked the question!
Good discussion. :)
I thought about putting it on a seperate partition, like Wvoice has, but
couldn't figure out why there would be a performance benefit. :)
On Thu, 2004-01-01 at 22:18, Jeremy Jackson wrote:
> Wvoice wrote:
> > Jerry,
> >
> > If you wish to avoid the recovery method, you can still use an external
log
> > with your RAID. If you're using hardware RAID, where the array spans
the
> > entire drive, create two partitions on the array (sda1 and sda2, where
sda2
> > is used for the log and is less than 256MB). If you're using software
raid,
> > then it's just as easy. Just create a couple of partitions on each
drive in
> > the array. You'll then create an md1 stripe across one set of
partitions
> > for data and another stripe (md2) on another set of partitions for
metadata.
> >
> > You can then use the following: mkfs.xfs -l logdev=/dev/sda2,size=32000b
> > /dev/sda1. You can substitute md1 and md2 accordingly.
> >
> > It won't matter that you're creating a metadata device and a data device
on
> > the same physical set of drives. You're still creating a new page
buffer
> > for the new logical device and this is where you'll speed up
performance.
> > The performance hit of physical disk seeks should be negligible. I've
> > actually benchmarked the difference between putting the journal on other
> > RAID stripes, other raids, fast disks, NVRAM, etc. You really shouldn't
> > notice much difference. Plus, placing the log on another stripe is
pretty
> > cheap to do.
>
>
> How did you test this... lots of metadata writes? I can't believe this
> is a good idea for performance.
>
> Regards,
>
> Jeremy
--
Jerry Haltom <jhaltom@xxxxxxxxxxxxxxxxxxx>
Feedback Plus, Inc.
raid_partitions.xls
Description: MS-Excel spreadsheet
|