RAID controller question...

[Stanley, Jeremy]

Compared to the cost of a commercial data warehouse, buying a COTS array
controller and a pre-built array of 5x4Gb LVDS-3 drives (which are
supposed to achieve double the transfer rate of single-ended SCSI) for
$2K is far cheaper (like around a 90% discount).

My interest here is akin to a drive system equivalent of the cluster
computer idea.  You can plug as many homogenous drives as you like into
the array (within reason) along with striping for speed and maybe parity
for redundancy, error-checking RAM in a large cache, etc.  The node it's
connected to, however, sees a single SCSI drive.  It's the JBOD (just a
bunch of drives) concept at the root of everything.  A lot of drives
that look like one, and spread the requested tasks out as evenly as
possible among one another.

And if you want more space, just plug in a few more drives and
rebuild...  Suddenly there's space for new partitions or to expand the
existing ones into.  Plus, most array enclosures have great cooling and
temperature management, while most controllers have built-in failure
detection to let you know, in some cases in advance, which drive is
failing.  Just my half-nybble.
-------------------------*-------------------------
Jeremy J. Stanley
Information Systems Administrator
Trend Community Mental Health Services


> ----------
> From: 	Mark Hahn[SMTP:hahn@coffee.psychology.mcmaster.ca]
> Sent: 	Wednesday, September 09, 1998 6:24 PM
> To: 	beowulf@cesdis1.gsfc.nasa.gov; extreme-linux@acl.lanl.gov
> Subject: 	RE: RAID controller question...
> 
> > Thanks for the lead...  Insight had the DPT PM3334UDW for $1275
> which
> > puts it in the same price range as the ICP GDT6518RD.  As for Mylex,
> 
> phew!  I'm not trying to belittle those of you who believe in RAID
> hardware, but these prices are enough to stop and make you think.
> $1275 is 50G of fast, cool UDMA disk.  yes, such disks will only
> sustain
> 12-14 MB/s apiece, but I've heard people report good results using the
> MD 
> driver under Linux.
> 
> remember, the touchstone of Beowulf is off-the-shelf, commodity
> hardware.
> $1275 might not blow the budget of a large cluster project, but it's 
> worth thinking about whether you'd be better off just using comodity
> hardware and mirroring, etc.
> 
> (and yes, before you ask, I _have_ measured 12-14 MB/s, sustained,
> Ext2 with 4K blocks, on run-of-the-mill Triton/Promise/ViaMVP3-type
> UDMA controllers.  depending on the processor, the bandwidth incurrs
> 7-20% CPU overhead, most probably due to the filesystem, not the disk
> interface.)
> 
> regards, mark hahn.
>