[Pc_Support] Re: MEdia Driven Universal Storage Array (MEDUSA)

[Bryan J. Smith]

On Sun, 2005-01-23 at 12:09, Jason Boxman wrote:
> Nice -- 5yr warranty.

Yes and no.  Yes the drives are ST3200822A and supposedly have 5 year
warranties.  But the box says 1 year.

> Ah, not RAID 5.  Sweet.

Not even RAID-3 or 4, or even traditional RAID-0 for that matter.

It's a really wild variant of what RAID-2, only strictly for ATA,
although there is nothing stopping it from being applied to SCSI
(although SCSI is buffered I/O).

ATA and WideSCSI are 16-bit data.  You make a 1:1 peripheral to bus
interface.  You do not buffer or block anything, you go _direct_ --
except for the parity write (which is not seen by the peripheral or
bus).  Two drives for 32-bit, four drives for 64-bit, plus parity disk.

Reads not RAID-0, which is striped blocks (e.g., 32KB from each disk),
but directly 32-bit or 64-bit wide from _all_ disks.  That's why it's
only ideal for a few I/O operations (like desktops or basic
appliances).  Writes are the same way, except the ASIC intercepts them
and calculates XOR, then writes to the 3rd or 5th disk which is parity. 
It seems work in real-time quite well.

If a drive fails, the other bytes are still read from the remaining
disks, but the ASIC also pulls from the 3rd or 5th disk that is parity,
and XORs out the missing disk and sends it in real-time.  Tom's showed
that this was much, much, much faster than even 3Ware's approach for a
failed RAID-5 disk.

> But if you did want 1000Mbps, you would need to buy a mainboard with yet 
> another PCI bus to segment the NIC?  (i.e. you would now need 3.)

Yes and no.

Yes, it would be ideal.  If I get a cheap 1000Mbps NIC that ties up the
bus, then I'm killing my disk I/O.

Yes, the NetCell SR5000 is going to burst up to 200MBps, and saturate
that bus.  So that's another reason.

No, if I get a _good_ 1000Mbps NIC, then maybe I won't need to.  266MBps
might be enough.

No, if I get a S2466 with 64-bit x 66MHz (533MBps), that might be much
better and enough to handle both -- assuming, of course, I get a _good_
1000Mbps NIC.

It all depends.

> May I ask why?  The NetCell card is at least $250.

Found one for $220, $230 shipped.  Right now I'm buying it for my
existing workstation.  Then I'll move it to the new MEDUSA system when I
upgrade my workstation.

I'll probably go with a SerialATA RAID controller, or maybe do software
SerialATA RAID, using the point-to-point SerialATA channels that are on
the dedicated 250MBps PCIe x1 channel of the nForce4's MCP.

> You can't get a hardware encoder too? ;)

First off, this is ATSC, _not_ NTSC.
NTSC is analog and unencoded.
ATSC is digital and encoded at 19.2MBps.
The common NTSC digital encoding is DV, 3.6MBps -- _huge_ difference.

I'll capture and use ATSC.
In fact, this is basically what the $1K DirectTiVOs do (only they use
QAM instead of the ATSC digital format).

ATSC cards already cost $200-300 (the pcHDTV is $189).
To buy a real-time ATSC hardware encoder, we're talking serious $$$
(over $1K last time I checked, but they may have come down since).
If I want to re-encode to a smaller format, I'll consider it,
_off-line_.

Which is why I want the extremely fast disk array.  ;->

Plus I have _yet_ to see one that works with Linux, or guarantees to
_never_ honor the broadcast flag.  If you can suggest otherwise, please
do so.  Otherwise I'm not spending $1K+ on something might not support
Linux, much less

> You expect you might fully utilize your 266MBps bus with that RAID card?

Yes.  RAID-XL at reading bursts in excess of 200MBps, and over 130MBps
when writing.

> I'd love to see some benchmarks whenever you get the box up.

Of course.  I'll have some on my existing workstation too.

> Yeah, I read about many hardware mods people were trying to avoid nuking their 
> S2460 boards.  The threads I read indicate it was still an issue even with 
> good power supplies.  Something about resistance on the +5V of the ATX 
> connector itself?

Trying to pump 70A through the single +3.3V and +5V lines is ludicrous. 
The new single processor Pentium 4s are the absolute worst, over 20W
higher than any Athlon, much less the Prescott Xeon 2P where you're
talking over 200A (not even ATX 2.0 will work for them)!

AMD, on the other hand, sells a 90nm 2GHz Opteron 246E with 4x the L2
cache that runs 10W _cooler_ (55W) than my 130nm 2GHz Athlon MP2400+
(65W).

> People were soldering an extra Molex on.

Yep.  Although then you have to wonder how they are dealing with other
details, let alone the soldering job.

> I'd be curious, but they go for >$200 on Ebay which is probably way outside my 
> price range when a good PSU will be another $100+.

I'm interested in selling the mainboard-only.  It's great for single
processor, plus I ran my Athlon 1.4GHz on the old, stock Antec 400W
(this is before the "True Power" series).  I eventually upgraded to an
Enermax 435W after adding a GeForce3 Ti200, and then to my current Antec
True Power 550W after I added a GeForce4 Ti4200.

> I imagine you mean just the mainboard, so it ought to be a bit less
> than $200.

I was thinking $70.

> Would there be any point in replacing my ServerWorks IIILE board with this?  
> It still only has two busses.

Correct.  Two busses (with the second bridged from the first), and one
memory.  Of course it has 2x the memory DTR (DDR), and 4x the CPU DTR (2
independent DDR FSBs, 2x if you use only 1 CPU).  But, again, I don't
recommend you put (2) newer Athlon MPs in it, older Thunderbird/Duron
ones, or just a single Athlon XP.

I'm still debating whether to keep it for my wife (a single Athlon
XP2600+), or sell it, if I go to a Tyan S2466 Tiger MPX at all.  I think
I might just stick with it for awhile and upgrade to the S2466 if I get
a bottleneck.

-- 
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.