Skip site navigation (1)Skip section navigation (2)
Date:      Sat, 10 Feb 2001 17:34:01 -0600
From:      Mike Meyer <mwm@mired.org>
To:        "Ted Mittelstaedt" <tedm@toybox.placo.com>
Cc:        questions@freebsd.org
Subject:   RE: Problems installing 4.x on large disks
Message-ID:  <14981.53225.50061.220090@guru.mired.org>
In-Reply-To: <001101c093b5$cfa47880$1401a8c0@tedm.placo.com>
References:  <14981.31829.637358.574815@guru.mired.org> <001101c093b5$cfa47880$1401a8c0@tedm.placo.com>

next in thread | previous in thread | raw e-mail | index | archive | help
Ted Mittelstaedt <tedm@toybox.placo.com> types:
> > Ted Mittelstaedt <tedm@toybox.placo.com> types:
> > >   Anyway, pardon me for being bigoted, but IDE disks have no
> > > place in a UNIX workstation.  With supported PCI SCSI adapters
> > > as cheap as they are, I for one fail to see why people are
> > > willing to trade away gobs of performance for a few extra
> > > gigs of hard disk space.
> > Because they *aren't* trading "gobs of performance" for a few extra
> > gig of hard disk space. While in the past, IDE drives pretty much
> > sucked (they were slow and ate CPU), if you get modern UDMA drives,
> > they'll keep up with all but the fastest SCSI drives in a
> > drive-to-drive comparison.
> true, only drive-to-drive.  But, to go to the next performance level
> you really need to get more spindles and aggregate them together with
> striping (ccd) or RAID (vinum)

Yup. And the last price/performance study I saw for that came down in
favor of a custom RAID box that had a bunch of UDMA drives in it and a
SCSI port to talk to the system. UDMA controllers are cheap enough you
can throw one in with the HDA and still beat SCSI prices.

Of course, you're not talking about a typical Unix workstation at that
point, either.

> > Of course, SCSI *is* a much better protocol. Two SCSI disks on one
> > controller take up one IRQ, and will perform much better than two IDE
> > disks on one controller, and slightly better than two IDE disks on two
> > controllers - which takes up two IRQs.
> > And it's more than just "a few extra gig". 50G scsi drives are around
> > $500. 60G UDMA drives are under $200. You're paying 2x to 4x more per
> > gigabyte for SCSI than IDE - and you only get extra performance in
> > multi-drive systems.
> And boy the quality of those $200 drives is right up there with the
> $500 ones - NOT!

Well, I haven't checked those. The SCSI drives were Seagates; the UDMA
drives were Maxtors, and those happen to be the drives I did my
testing on. The Maxtor was slightly faster than the Seagate once you
got the UDMA stuff turned on.

> The wide price differences of SCSI is only present in the largest
> drives.  If you really want 50GB your speediest performance is to
> take 3 20GB SCSI disks and stripe them.

Yeah - obsolete drives tend to bottom out in price. The last time I
bought a SCSI drive (about a year ago), the 10GB SCSI drives cost
slightly more than 2x what the 10GB IDE drives did. That's no longer
true, of course. On the other hand, to get the best performance, you
don't buy obslete drives.

> The other thing is that with multiple overlapping seeks a SCSI
> controller and driver is far more efficient about making decisions
> about how to best move the head around.  With a regular Adaptec 2940,
> for example, I think you can stack up to 255 commands into the adapter
> processor and the adapter can chew on them while the main CPU goes and does
> something else.

Well, if I spent more on the controller than I did on the drives, I'd
certainly expect it to contribute something to the performance. On the
other hand, the same $100 spent on another hundred or so megabytes of
RAM will mean most workstation users don't page, which is liable to
make even more difference.

> > Bottom line: if you only have one drive, the extra cost of a SCSI
> > drive would be better put into more RAM. For low-end servers, I buy
> say rather "if the drive has to be as large as possible" and I'll agree
> with you.

Nope. My tests were on 10MB IDE & SCSI drives, and the IDE drive
delivered better throughput with less CPU load at a lower price. If
you say rather "if cost is no object", I'll agree with you.

Again, if you don't believe my numbers, run your own tests and tell us
about it.

	<mike
--
Mike Meyer <mwm@mired.org>			http://www.mired.org/home/mwm/
Independent WWW/Perforce/FreeBSD/Unix consultant, email for more information.


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?14981.53225.50061.220090>