Consumer and Enterprise ATA, plus 3Ware and NetCell -- WAS: LinuxVIA Chipsets

["Bryan J. Smith" <b.j.smith@ieee.org>]

On Wed, 2005-01-05 at 16:28, Steven Pritchard wrote:
> Not so recent example:  In 1998, I started buying IBM hard drives
> after hearing nothing but good things about them for around a year.
>  ...
> So, after all that, many of you might recall the 2001 Deathstar issue,
> where just about every IBM desktop drive ate itself after 3-6 months.

Actually, the problem was putting 5-platters in a 1" form-factor.
If you had the 3-platter versions, they lasted much, much longer.

In a nutshell, the problem with ATA heat and how they are used.  Most
ATA goes _inadequately_cooled_ in most enclosures.  And ater the
Deskstar issue, IBM went to stating their drives are _only_ for 14x5
operation.  Many other vendors have also followed suit.

Some vendors now have different lines -- one for 14x5 "consumer" ATA,
another for 24x7 "enterprise" ATA, sometimes the latter is the same as
their SCSI line, only with ATA logic.

Some vendors who only do ATA, like Maxtor, have gone to new tolerance
testing.  Those drives that test to high tolerances are labeled 24x7
"enterprise" ATA, and the others are shipped out as 14x5 "consumer" ATA.

Finally, the new "golden rule" on ATA is the 40C rule.  If an ATA drive
exceeds 40C, it's lifetime is drastically reduced.  Just a consideration
these days, since 7,200rpm is found everywhere.  If it isn't cooled, 40C
is a realy problem.

> To say I got bit by that would be a monumental understatement...  I
> had just started to build large-ish RAID boxes with 3ware cards around
> that time.  In addition, the primary drive in my desktop was from
> around that era (it died in January of 2003), and I have a drive from
> around then that's been threatening to fail on me for a while now in
> one of my servers (one of only two that is old enough that I don't
> have mirrored drives in it, of course).

I don't build even a desktop these days without at least a 3Ware 7006-2
or 8006-2.  And I prefer 3Ware Escalade 8506-4, 8 and 12 for servers. 
The 9500S-8 and 12 is also a consideration, especially as the Linux
support matures.  Even Tom's recently found that even 500MHz XScale
microcontroller-based SATA and U320 cards still can't compete with
3Ware's ASIC at both I/O queuing as well as transfer rates, especially
during a rebuild.

With that said, the new company that is really interesting for the
_desktop_ (not workstation/server) is NetCell and their "SyncRAID." 
They have what they call "RAID-XL" which is kinda like RAID-3/4,
dedicated parity disk, but then they use "wide" writes instead of
stripped blocks.

I.e., instead of striping a disk every 32KB or so, with an
incoming/outgoing 16-bit ATA stream from/to an ASIC or microcontroller,
they read/write a _direct_ 32-bit or 64-bit PCI stream to either 2 ATA
discs or 4 ATA discs, respectively.  In the case of a write, on-board
ASIC+DRAM then buffers the data for the parity disc, which is written
later.

What is surprising is not only the read performance, which is even
better than a 2 or 4-disc RAID-0 (because all discs are use
simultaneously), but the write and disk lost performance too!  The ASIC
engine can "keep up" well.

Kicker #1 is that you must use _exactly_ 3 discs (2 data, 1 parity) for
32-bit PCI, and 5 discs (4 data, 1 parity) for 64-bit PCI.  But this is
a very ideal configuration.

Kicker #2 is that this is clearly a _desktop_ operation with only a few
commands being queued.  On a server, direct disc access loses its
advantage as lots of random accesses are happening all over the place.

But because it's oriented towards the desktop, the final nice detail is
that NetCell did away with the SCSI interface, and makes the card look
like a single ATA drive with NCQ (native command queuing).  And with
just a small GPL patch (largely for ATA PCI ID), the Linux kernel has
full support.

The only reason I haven't bought one is because they still lack Linux
monitoring tools.  But they promise they will deliver them in the
future.

> Anyway, after that mess, IBM sold off its storage division to Hitachi.
> Since then, I've heard *nothing* bad about Hitachi drives (and a
> couple of good things, the usual stuff I always heard/said about IBM
> drives).  A couple of months ago, we were building a 2U rackmount with
> 12 250GB SATA drives, so we were faced with the choice of WD, Maxtor,
> or Hitachi.  (Seagate tops out at 200GB, and Samsung only goes up to
> 160GB.)  I *detest* WD drives, and I don't like Maxtor much better, so
> we went with Hitachi on that server.  It remains to be seen if that
> was a wise choice, but we're planning on sticking with Hitachi for
> large drives for now.

It's all about testing.  Maxtor and other "quantity" ATA vendors don't
test well, except for those drives that they label "enterprise" and sell
to storage array and select OEMs for servers.  Seagate has been
excellent for me with their Barracuda ATA series, even if they hold back
on the platters and, correspondingly, the size.

Fewer platters is good IMHO.  ;->

> I've softened up on my "no nForce" stance of a couple of years ago,
> but it still wouldn't be my first choice.  It seems that while
> nForce-based boards are supported well these days, they still aren't
> as stable as VIA-based boards for whatever reason.

I've heard differently.  The _only_ thing I've heard that "isn't stable"
is the GPL forcedeth driver.  But when it comes to systems, I almost
_never_ use the on-chipset NIC for corporate systems.  Just can't trust
their variations.

IWill sells a nForce3+AMD8131 dual Socket-940 in the DK8N that is most
excellent.  Their forthcoming nForce4+AMD8131 DK8E will probably be the
one I will buy.

> Since VIA-based boards tend to be significantly less expensive, I
> still recommend them (even for the Windows users who come into our
> store).
> I would definitely stay far, far away from any of the other chipset
> brands right now, if for no other reason than they are such a small
> percentage of the market that they are more likely to have bugs and
> less likely to get quick fixes when there are bugs.

SiS is fine for single CPU desktops with basic configurations.  I've had
issues with 3Ware cards which use an internal bridge, their PCI I2C/DMA
doesn't seem to like them.

Intel has largely _tanked_ since the i440+PIIX series, _except_ for
their i7500 series which is licensed from ServerWorks.  I've heard their
i7200 series is also fairly good.

As I mentioned before, the i875 is the same chip as the i865, but tested
to higher tolerances.  Unfortunately, many mainboard vendors reused i875
designs for i865 chips and found massive reliability issues.  Asus was
one such vendor.


-- 
Bryan J. Smith                                    b.j.smith@ieee.org 
-------------------------------------------------------------------- 
Subtotal Cost of Ownership (SCO) for Windows being less than Linux
Total Cost of Ownership (TCO) assumes experts for the former, costly
retraining for the latter, omitted "software assurance" costs in 
compatible desktop OS/apps for the former, no free/legacy reuse for
latter, and no basic security, patch or downtime comparison at all.




-
To unsubscribe, send email to majordomo@silug.org with
"unsubscribe silug-discuss" in the body.