Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 9 Mar 2000 07:24:07 -0800
From:      Scott Hess <scott@avantgo.com>
To:        Greg Lehey <grog@lemis.com>
Cc:        Scott Hess <scott@AvantGo.com>, james@icorp.net, freebsd-questions@FreeBSD.ORG
Subject:   Re: RAID/config questions
Message-ID:  <20000309072406.A10953@river.avantgo.com>
In-Reply-To: <20000309141407.L58942@freebie.lemis.com>
References:  <38C45697.D736070F@icorp.net> <20000306183129.B2525@river.avantgo.com> <20000309141407.L58942@freebie.lemis.com>

next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, Mar 09, 2000 at 02:14:08PM +1030, Greg Lehey wrote:
> On Monday,  6 March 2000 at 18:31:29 -0800, Scott Hess wrote:
> > On Mon, Mar 06, 2000 at 07:08:39PM -0600, James wrote:
> >> 2.  Which will give me better performance?  RAID or RAID5?  I
> >> know RAID5 will give me more disk space, but is there any significant
> >> i/o performance cost?
> >
> > My most recent experience was with an external SCSI-SCSI RAID
> > controller attached to an Adaptec differential host adapter on
> > FreeBSD3.3.  With six drives plus a hot-spare, I found that
> > RAID5 (5 drives plus parity) and RAID1+0 (3 drives worth of
> > mirrorred and striped data), the performance was pretty similar.
> > The deciding factor was that when we pulled a drive, rebuilding
> > under RAID5 really degraded performance (it could carry about
> > 1/3 the tps relative to when it wasn't degraded), while rebuilding
> > under RAID1+0 was only marginally noticable.
> 
> I'd guess you didn't do any particularly rigorous performance testing.

Rather than one-line sniping, could you elaborate?  Performance
doesn't degrade when rebuilding the hot spare into the RAID5
array?  RAID1+0 isn't comparable in performance to RAID5 for
some tasks?  Rebuilding wasn't marginally noticable on the RAID1+0
array?  If you have specific quibbles with my statement, I'd be
more than happy to go go back and do some different tests, if it
looks like we missed something.

Note that I didn't say that RAID1+0 was better than RAID5 for all
possible uses, nor that RAID5 became significantly degraded for all
uses, or anything at all about vinum.  The above _did_ happen for the
home-grown test suite we were using to replicate our production I/O
usage.

> >> And generally speaking, how much of a performance degredation may I
> >> see (if any) in going from a non-raid SCSI to a RAID or RAID 5 setup?
> >
> > Hardware RAID shouldn't degrade performance at all.
> 
> I suppose that depends on what you mean by "hardware RAID".  With
> conventional RAID controllers, the difference is really where the
> software runs.  In "software RAID" it runs in the main CPU, usually
> quite a powerful processor.  On RAID board the processor is usually
> much slower.  On top of that, RAID-5 requires many more I/O accesses
> for a write than RAID-1 does.  This slows down writes, and there's not
> much you can do about it.

My assumption is that if you're going to spend the extra money
to purchase a RAID5 controller, you're going to spend enough
to purchase one that isn't doing parity calculations on a 6502.
RAID5 has to touch all the data, but the calculations themselves
are not rocket science - the CPU on the controller only has to
stay ahead of the drives it's writing to.

In general, I agree that you're better off with shared access to
a powerful (fast or wide or nearby) resource than dedicated access
to a less powerful resource.  But unless the controller cannot
stream through memory faster than data can be streamed to the
disk, it shouldn't make a difference.

Personally, I'd never bother going with hardware RAID unless I
intended to put a serious amount of disk space on the system in
the first place, at which point it makes more sense to buy a
decent controller.

> > Software RAID0 or RAID1 shouldn't have much impact, RAID5 can have
> > more because it has to actually muck with the data.
> 
> Software RAID doesn't pose much of a load on normal modern CPUs.

Much of the time you're using RAID because disk I/O is the bottleneck,
meaning that the CPU load is besides the point in the first place!
At least where I've had to use RAID, I'd have gladly raised the I/O
CPU load 10x if that doubled the disk I/Os per second that I could
do.

Later,
scott


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20000309072406.A10953>