[SunRescue] Question about SSAs

[Jonathan Katz]

David wrote:
> at the problem and use very narrow, very wide stripes at the outer edge of
> the disks. I generally work with an application that has an RDBMS component

When I was working on deploying the last E10K I worked on at a client site
(a lot of domains running SYBASE of all things) this is exactly what we did
on the Symmetrix. The theory is called 'wide thin disk striping' and is
covered in the following Sun Blueprints document:

http://www.sun.com/blueprints/1000/layout.pdf

> this is where you get the magic of the Symmetrix-- it has such a HUGE cache
> (in this case 8GB) it can afford to delay writes for a long time so it can
> re-order them to reduce or eliminate contention between elements on the
> disk. It's beautiful, really.

Yes, the problem with here is that it sometimes takes some arguing with
the EMC PS/SSE-type folks to layout the Symmetrix with teeny-tiny
'hyperdrives.' The SAs I worked with in deploying that E10K took 2-3
days of arguing w/ the EMC guys to get them to do 2Gig instead of the more
standard 8-16GB drives. A few of the EMC guys didn't understand why we
were doing what we were doing, either :(

The problem with the EMC cache on a Connectrix is that it's a bottleneck--
you'll have a dozen systems feeding into it from a dozen controllers, and
there is ONE cache-- it may be HUGE but there is still only ONE cache. IIRC
this cache isn't mirrored or battery protected, either, and can provide for
a single-point-of-failure. *points to battery backed-up ECC cache on RSM2000,
A3500 and T3*
 
> So in short it's easy to use the entire disk on an array like the Symmetrix,
> but a bit trickier on JBOD like A5x00. Or the T3s or A3500s for that
> matter... neither of them are as smart as the Symmetrix.

The Symmetrix only does hardware mirroring, you still need to use veritas or
disksuite to stripe the 'hyperdrives' that the EMC carves out. It takes a
buttload of tweaking to get the hyperdrives to represent themselves as
the teeny, tiny 2GB chunks you want to stripe. In both cases it's a lot of
administration-- it's just different as what you have to administrate. Ever
have to have a junior admin run 'format' a few hundred times? (otherwise
each EMC hyperdrive won't have a disklabel and that causes a 'WARNING' 
message in /var/adm/message.... and apps like Tivoli, BMC Patrol, etc will
go batshit if they see 'warning'.)
 
> The thing about the A5x00 series is that since they're 'dumb', they don't
> suffer from the reliability problems that you see with Sun's attempts to do
> hardware RAID... and they're so (relatively) cheap and have such great rack
> density you can literally just throw tons of 9GB or 18GB spindles at
> something.

The A3500 was a disaster of a product. Only now, when everyone is sick of
dealing with it with proper mangement software (RM6.22) and the latest
firmware is it 'useful.' The same deal with my RSM2000 and even that has
some quirks (most of that is because I don't have a 240V UPS at my schwank
pad.) Sun's latest hardware raid products only do raid 5 I haven't worked
with a T3 hands-on, but that's what I gather.) You can have them stripe
then do mirroring in software. They also have cool stuff like partner
pairs, etc.

The 5x00s are just FC->SCSI JBODs much like the EMC just w/o the hardware
mirroring. When you mirror your disks in an A5x00 do you mirror to drives
w/in a single JBOD or to another JBOD? I've seen debates both ways...
Software raid 5 is just sloooooow. Software mirroring is OK over multiple
controllers.

It's a gorgeous Sunday. I'm not spending the rest of it at /dev/console.

Take care kids!

-Jon
--
Jonathan Katz [] jon at jonworld.com [] http://jonworld.com
Cell/Pager: 317-698-4023 aka 3176984023 at NO=SPAM.messaging.sprintpcs.com