[NTLUG:Discuss] RE: true hardware/intelligent ATA RAID -- XScale SATA RAID solutions arrive (including PCIe)

[Bryan J. Smith]

[ Reviving an old thread ... which I can now do so since XScale-based
(yeah, no more "10-year-old, slouchy i960/IOP30x"!) SATA RAID solutions
_are_ now available! ]

On Wed, 2004-07-21 at 12:58, Bryan J. Smith wrote:
> Okay, did a bit of reseach:  
> http://developer.intel.com/design/iio/index.htm?iid=ipp_embed+proc_iop  
> As I mentioned earlier, the 8030x is i960-based.  You're talking a
> single integer pipe microcontroller.  Pretty old, pretty crappy for
> high-speed disk I/O.  Even by Intel's own marketing, you can only expect
> a 25% performance increase with the "last generation" 80303.
>   ... cut ...
> The newer XScale series of products are pretty modular.  They will all
> be marked 8031x or higher.  The latest support PCI-X 1.0 bus arbitration
> upto 64-bit x 133MHz, including on-board communication I/O (e.g., GbE). 
> The XScale has upto 7-issue integer pipeline design, 32-bit ARM ISA with
> 64-bit data bus -- blows away traditional microprocessor design at I/O. 
>  ... 
> Unfortunately, all LSI ATA RAID cards are only i80302 or i960.

- The I/O Processor Bottleneck Removed

As noted earlier, Intel's aging, non-superscalar i960 and subsequent
8030x/IOP30x series of I/O Processors (IOP) were rather weak and
limiting in performance.  The single-piped, 32-bit 100MHz flavors were
barely capable of pushing maybe 66MBps maximum.  But that's all that was
available in "true hardware RAID" for ATA, other than 3Ware's
ASIC+SRAM-based approach in its Escalade series (which I preferred).

The superscalar Digital StrongARM 110 206MHz has been in use in
multi-channel Ultra80 (and faster) SCSI controllers since the very late
'90s.  It was 5x faster at I/O than even the "best" 80303 of the i960
series.  It's subsequent, 7-issue superscalar Intel XScale processor was
starting to show up in SCSI RAID solutions just a few years ago,
offering even greater performance.

And now we are finally starting to see the first SerialATA RAID
solutions with it too!

- LSI MegaRAID SATA 300-8X, TekRam 1110/1120, 1210/1220

LSI Logic and TekRam have recently introduced SATA RAID produced based
on the XScale.  The two IOPs you will see are:  

- IOP331 (i80331) "Lindsay" 133MHz PCI-X 1.0 (1GBps)
- IOP332 (i80332) "Dobson" 133MHz PCI-X w/8-channel PCI-Express bridge

Both are going to exceed the throughput capability of the older IOP30x
series by 10-fold or greater.  Marketing aside, the synchronous clocks
for CPU-I/O-memory interconnect is effectively 64-bit x SDR133
(1.06GBps) or DDR266 (2.1GBps), regardless of the core clock or memory
buses (which may be 333/667 memory/CPU).  This matches well to PCI-X 1.0
slots of 66-133MHz.

The LSI Logic MegaRAID SATA 300-8X is an 8-channel solution with the
IOP331.  It has a fixed configuration of 128MB of DDR333 ECC SDRAM. 
Battery backup is an option.  Both SATA/150 and SATA2/300 (300MBps) are
supported with the on-board controllers (one controller per 4 channels).

The TekRam solutions are similar.  The 1110/1210 (fka the SIR-804)
4/8-channel solutions also deploy the IOP331 with 128Mb of DDR333 ECC
SDRAM.  But TekRam is also offering a new variant in the 1210/1220
series of 4/8-channel solutions with the IOP332 for PCI-Express (PCIe)
x4 slots (assumingly backward compatible to PCIe x1).  TekRam also
offers RAID-6, two parity stripes, for 2 disc redundancy.  This is
important as ATA drives are typically tested to lower tolerances for
only 14x5 operation (not 24x7).

Pricing on the TekRam solutions start around $350 for the 4-channel,
1110 product and scale past $500 for the 8-channel solutions.  While
this is not commodity, it does compete well against XScale SCSI RAID
solutions -- especially when you factor in the cost of drives.

I would love to see a review pitting these new XScale microprocessor +
DDR-SDRAM "buffering" solutions versus 3Ware's ASIC + SRAM
"non-blocking" solutions.  The previous i960/IOP30x "yesteryear"
solutions didn't fair well, but now SATA RAID is becoming mainstream and
pushing more leading-edge technology adoption on-par with leading-edge
SCSI RAID.

- Consumer PCIe Product?  What about boards with both PCI-X and PCIe?

I'm still hoping someone introduces a budget-conscience, but true
hardware ATA RAID solution for PCIe x1 with 2-4 channels for sub-$250
(possibly sub-$150 for 2 channels).  Since PCIe x1 will be commonplace
in near-future mainboards, and a dedicated I/O channel for storage is
severely needed -- instead of taxing the shared PCI bus, or requiring a
not-so-commodity mainboard with PCI-X slots.

The cost of the IOP332 will make this prohibitive, so I hope 3Ware
designs a PCIe ASIC for a new Escalade series product geared towards
PCIe.  Until then, the commodity PCIe future looks to be filled with
"FRAID" software solutions with trick-firmware but are 100% OS
driver/logic driven.

Another idea I had is a card that "flips" to offer both PCI-X and PCIe
(you'd have to move the bracket, but that's it).  This has been done
before in the past.  The IOP332 is certainly capable of offering
jumper/card-selectable bridging/non-bridging of PCIe as appropriate.  It
would allow vendors to create a _single_ PCB and product solution, for
all interface offerings.


-- 
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.