🚀 go-pugleaf

In article <40pvso$s4c@abyss.West.Sun.COM>,
Bruce Adler [Contractor] <adler@West.Sun.COM> wrote:
>What, if any, is the difference between RAID 4 (as defined by RAB) and 
>"RAID 7"?  

I'm sure Willett will spew some marketroid buzzwords, signifying nothing.

As far as I can tell, the only real difference is that "RAID 7" has a 
very large cache in between the bus and the disks, which it uses to 
reorder writes to avoid the short write performance issues caused by the 
need to read-write-read the relevant blocks on the parity disk.

Because it manages to almost always write in stripe-size increments, it 
can probably give RAID 5-like performance, but with the easy 
expandability of any RAID 4 system -- you just add more disks; no need to 
worry about parity striping as you do with RAID 5.

This is all well and good, but not exactly revolutionary.  Log-structured 
filesystems or any other method of teaching the buffer cache on the host 
doing the writes to avoid short writes if at all possible -- and probably 
to write in multiples of the stripe size as well, which isn't really very 
hard -- will produce exactly the same effect with *any* RAID 4 array.  
All StorComp do, so far as anyone seems to be able to find out, is move 
the cache to the wrong end of the SCSI bus.  The wrong end, of course, 
because with it on the disk side of the bus the way they do it, you 
destroy the write-through semantics of the UNIX buffer cache for 
metadata.  This means that on a power hit, you can lose the whole 
filesystem, instead of just whatever was being written when the power 
went out.

The basic problem with calling "RAID 7" a new RAID level is that, well, 
it just isn't.  All it is is a RAID 4 with a big cache on the front.  
RAID levels are supposed to refer to the technology of the *disk array*, 
and the method by which data and parity is distributed amongst the 
component spindles.  If we want to play it StorComp's way, we need a new 
RAID level for every new caching scheme or write-reordering trick anyone 
ever tries -- I shudder to think how many RAID levels the various 
Berkeley grad students would have gone through. :-)  This opens us up to, 
in effect, "Gee, this disk array is RAID level 4 if I use it on my VAX 
which has an ancient dumb version of UFS which does 512-byte writes, but 
when I plug it into my RS/6000, which has a journaling file system, it 
becomes RAID level 77!".

Worse, under NetBSD, where I do much of my work, different SCSI 
controller drivers sometimes cause slightly different algorithms to be 
used to reorder _all_ writes, filesystem or no.  Does that mean that if I 
have two scsi controllers in my machine, with identical RAID arrays 
attached to them, one is RAID level 92 and one is RAID level 26?

Clearly, this is to be Considered Harmful.  StorComp's 
sleazy way of using the net for advertising doesn't help much either. :-)

Nor, in my book, does their complete lack of understanding of the way 
their cache architecture can lead to catastrophic filesystem damage.


-- 
Thor Lancelot Simon                                               tls@cloud9.net

Somewhere they're meeting on a pinhead, calling you an angel.


-- 
Thor Lancelot Simon                                               tls@cloud9.net

Somewhere they're meeting on a pinhead, calling you an angel.