[SunRescue] Excuse my gross oversimplification of RAID

Christopher Byrne rescue at sunhelp.org
Sat Dec 2 16:28:55 CST 2000


All,

Excuse my gross oversimplification of RAID issues in my explanation, I
figured it was long enough already, so I left out some of the less common
raid levels. A lot of you brought up some great points that I wanted to
address.

First of all excuse my brainfart on reversing 0 and 1. It was REALLY late
when I wrote that :-)


James Lockwood brought up a great point about there being heat problems in
the 12 bay multipack, and I agree that if you wanted to stick twelve of the
original 10k drives in there you would probably have a heat problem, never
mind the horrible noise involved. I believe that Sun never used more than
7200 RPM 18.4 gig disks in the 12 bay MultiPacks, but I might be wrong.
However newer 7200 and 10k drives are a lot cooler and quieter then they
used to be. There still might be heat issues, but with good drive selection
they could be minimized. Check out www.storagereview.com for more info on
drives.

I was not aware there was an impedance issue with U160 drives in the
multipack, however I will take his word for it. It's not hard to create an
impedance issue on an LVD SCSI chain, a slightly bad crimp in a connector
can do it easily, never mind an additional six drives.

I also agree with his price assessment. I have seen bare 12 bay MultiPacks
going for between $350 and $1500 with most falling into the $750-$850 range.
As to technical issues being the only reason why they removed the 12 bay
MultiPack from their product line I'll have to disagree. I worked with
several former Sun storage engineers who told me that they could have fixed
the technical issues if the company wanted to bother, but that with their
lower end SAN/NAS solutions coming in at the same capacity as the MultiPack
(and for WAY more money, albeit with more capability) they didn't want to
erode that market share.

The statement of raw disk capacity was just an illustration since most
people don't really fully grasp how much (or how little) a terabyte really
is. As Paul Gianocavo stated, in practice almost no-one would actually
create a raw storage array of that size without some type of redundancy. In
my case I would probably configure a 0+1 array to maximize performance while
still maintaining 100% redundancy. A properly configured 0+1 on a good RAID
controller is probably the fastest RAID solution, though I may be wrong.
With that solution I'd end up with 440gb useable.

I also agree with Mr. Lockwood in that if you need this level of storage you
are most likely best served by another approach, it's just fun to speculate
about having a TB array on your desk. Realisticaly for price/performance I
would probably stick 36gig U160 disks in (about $700 a pop), or screw it and
convert over to FC-AL. The great part about fibre channel is that the drives
are reativley cheap, it's the cables and the HBA that really kills you. An
FC-HBA can easily cost more than the computer it's attached to.

I do take issue with the exception Paul Theodoros made for my Raid 5
explanation in calling it striping with rotated parity. While this is true,
it is not generally referred to explicitly as having rotated parity. RAID 5
could also, and possibly more accurately referred to as having distributed
parity, because the decision of where to put the parity chunk is up to the
vendors implementation, and it is not necessarily always rotated, it may be
distributed across an array to optimize access patterns if an intelligent
RAID controller is being used.

RAID 3, 4, and 5 all offer striping with parity, however 3 and 4 both use
dedicated disks to do it. All three actually have advantages and
disadvantages, but RAID 5 seems to be the best balance between read
performance, write performance, and redundancy in a gernal purpose storage
environment. There are also vendor propietary solutions such as RAID-s in
which the parity calculation  is preformed in the firmware of the drives in
the RAID group.

Paul made some great points, and filled in the blanks of my explanation of
RAID levels. He is correct in that there are no official raid levels since
they are vendor controlled, however the original papers on RAID universally
defined levels 0 through 5 (some went further). And I also highly recommend
S/W expert as a source of information. In fact I stole the term E-JBOD from
there.

Also Gregory LeBlanc brought up the question of other names they have seen
the technology under, and it's a very valid question. The problem lies in
that vendors get to call their RAID levels whatever they want because there
is no enforcing standards body. Outside of the agreed upon levels of 0
through 5 a vendor can pretty much call their solution whatever they want,
and often do.

The most common rotation of names you see is vendors (and IT people for that
matter) calling a RAID 0+1 a RAID 10, or vice verse. There are distinct
differences between the two. In RAID 0+1 you mirror the entire stripe set.
This gives you 100% redundancy across the mirror, but this only allows for a
drive on one side of the mirror to fail. In RAID 10 you make a stripe set
out of mirrored pairs. This allows one drive from each pair to fail without
losing data, but if an entire mirrored pair dies you lose data. Overall RAID
10 is probably the better solution for redundancy, but it also has more
overhead because it generally requires volume management of the stripe set,
whereas 0+1 can generally be done entirely in hardware. In fact you even see
cheap IDE RAID controllers with 0+1 being built into motherborads now.

As I said the problem is that a vendor can call their RAID whatever they
want, and many vendors who are actually using RAID 0+1 are calling it RAID
10 or the other way around. I've even heard some vendors say that they were
interchangeable, or rather a direct quote from a sales rep "Oh thet're the
exact same thing"

Once you get outside the 1-5/0+1/10 area RAID naming is a total free for
all. The industry hasn't really settled down on what to call mirroring a
stripe set with parity or srtiping with parity of mirrors. 50, 51, 100
etc... al get tossed around, you had better ask the vendor exactly what they
mean if they are talking about anything other then 0-5

Chris Byrne
=======================================
The eyes may be the windows on the soul
But the word is the doorway to the mind
=======================================





More information about the rescue mailing list