[NTLUG:Discuss] NetCell may give 3Ware a run for their money on the desktop/entry-level server ...

[Bryan J. Smith]

The NetCell products look extremely innovative:  
  http://www.netcell.com/  

In addition to offering standard RAID-0 and RAID-1, they offer what they
call RAID-XL.  It's really a variant of RAID-3/4, where you use a
dedicated parity disk, instead of a striped parity slice across all
disks.

Where traditional uC+DRAM RAID uses latent buffering, and 3Ware
ASIC+SRAM uses non-block switching, NetCell takes it to a new level.  By
mandating a "fixed" configuration, it can offer the most _optimal_
balance of performance and cost.  In fact, I'm surprised someone hasn't
consider this before -- as most RAID-3/4/5 implementations are typically
3 or 5-disc at the "entry-level" end anyway (as well as the fact that
3-disc 2U and 5-disc 3U hot-swap bays are available).

NetCell, when doing this "RAID-XL", uses a direct 2x16-bit (32-bit PCI)
and 4x16-bit (64-bit PCI) access to 2 or 4 discs, respectively.  The 3rd
or 5th disk is then the parity disc.  On a READ, the data comes
_directly_ off of each disk -- ATA is read in 16-bit words.  Again,
2x16-bit for the 32-bit PCI version, 4x16-bit for the 64-bit PCI
version.  By mandating this "fixed" configuration, the ultimate in
non-blocking I/O directly to the end-user ATA device can be achieved.

[ SIDE NOTE:  I wonder how it works for PCI-Express?  I think
PCI-Express is still a logical 32-bit or 64-bit width, encoded for each
1-bit, bi-directional channel?  I'll have to verify. ]

On a WRITE, this still occurs too, but the 64-512MB SDRAM buffer is then
used with the built-in XOR engine to calculate the parity -- which is
written to the dedicated parity disc.  Again, the dedicated parity disc
is nothing new, as Promise promotes RAID-3 for desktops, and NetApp
promotes RAID-4 for its filers (commonly implemented as NFS/SMB NAS-like
devices).  Most cards seem to offer 64-128MB of SDRAM, which is enough
to buffer most typical writes on a desktop system, possibly a small
departmental server.

And instead of creating a dumb block device driver that ties into a
typical SCSI subsystem of the OS, NetCell provides a standard ATA
driver.  Now this means it suffers the limitations of ATA (e.g., if your
OS doesn't support LBA48 addressing, you'll have issues --
NT4SP4+/2000SP1+ required), but it does allow inherent support of the
controller with small modifications of existing ATA drivers.  This
includes their current "beta" patches to Hendrick's ATA code in the
Linux kernel (and can be GPL I assume?).

Given the sound approach to technology, I have only 2 remaining
questions:  
1.  How does it queue operations, if at all?
2.  How does one manage the unit?

Regarding #1, I have to assume the on-board ASIC an offer additional
capabilities.  And the SDRAM does probably buffer reads/writes as
necessary.  And new ATA devices approaches like Native Command Queuing
(NCQ) removes the benefits of traditional SCSI or intelligent ATA RAID
cards like the 3Ware Escalade with its ASIC or new ATA RAID cards with
XScale superscalar microcontrollers.  According to Tom's Hardware Guide
benchmarks (link below), this seems to be one of the major limitations
of the approach, which would limit its potential for servers.

Regarding #2, there are management utilities for Windows.  For Linux,
there are none (yet?).  So I might not be changing away from 3Ware
immediately for servers, but for desktops, I might start prototyping. 
Especially if they introduce a PCI-Express version (which brings us back
to the 32/64-bit "logical organization" question above).  A 3-disc
RAID-XL looks like an idea solution for desktops, performance and
redundancy in the age of largely 1-year warranty ATA drives.

I'd still like to see some more comparisons of its RAID-XL writes versus
other solutions like the 3Ware Escalade 9500S-8 and LSI Logic MegaRAID
SATA 300-8X, which cost twice as much.  Tom's Hardware showed largely
only _read_ benchmarks.  Although the real-time XOR calculation from
parity _was_ impressive when a disk is lost -- that I have to admit. 
Tom's conclusions were pretty good, talking about both the "benefits"
and the "limitations" of the approach:  
  http://www.tomshardware.com/storage/20031128/netcell-14.html  

I will watch NetCell closely.

-- 
Bryan J. Smith                                    b.j.smith at 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.