Re: fsck is dead

jplooney@compapp.dcu.ie
Tue, 6 Jul 1999 20:01:40 +0100


On Fri, Jul 02, 1999 at 09:46:00PM +0200, Kai Henningsen mentioned:
> lindahl@cs.virginia.edu (Greg Lindahl) wrote on 29.06.99 in <199906292321.TAA09459@adder.cs.Virginia.EDU>:
> > HA is half of the key. Fast fsck is the other half. You can't avoid
> > the fact that the OS scribbling on your dual-ported disks is a single
> > point of failure. After that, do you really want to bring up a backup
> > computer without an fsck?
> So isn't that obviously NOT the right way to do it?
> What would you do if that disk broke?
> For situations like that, you want completely separate hardware. On
> separate UPSen. On separate power lines. Probably in different cities.
> (California-style earth quakes. Tchernobyl-style nuclear plant blowups.
> Australia-style power cable burnouts.)
> Dual-ported disks is not a desaster plan.
> (I remember hearing of a desaster plan of a big manufacturer - making hard
> disks or something like that - they supposedly had a complete second setup
> at the other side of the city, and the plan said "if something takes out
> both places, we're going to have more serious trouble than caring about
> hard disk production".)

There are commerically availible boxes that do that sort of thing for you.
A fully kitted out Sun Microsystems A7000 disk array can take 40TB of
disks, and RAID them anyway you want. You can then have up to 40 SCSI
connectors (or a dose of fibrechannel connectors, mainframe, NFS -
whatever) that can connect the disk array to all the machines you can fit
into it; the A7000 is a fairly beefy unix machine (I think it's running
umacs for the 88000 series), that makes it's end of the SCSI bus look like
a load of disks.

It has 8 CPUs that will do all the RAID for you, as fast as you can throw
data at it (and with 40 UWSCSI connectors, that's a lot). The 8 CPUs are
setup as 4-way SMP nodes; if one CPU dies, the A7000 issues a SCSI retry,
and the other set of CPUs deals with the request. The cache memory is
"reflexive", in that any data written to the A7000 gets written to two
separate banks of parity memory - so if there are errors that parity can't
fix - doesn't matter - there are two copies of it.

For the real paranoid, you can hook it upto a second A7000 via fiber, so
that it mirrors everything that's written to it over the fibrelink
something like 20 miles away, and it doesn't issue a "that's to disk"
before it's in the cache of both A7000s (which being mirrored, is as good
as being on disk). So, if you take a hit from a nuke, it's cool - no data
loss, as it's all at your secret fallback site. I don't know what these
boxes cost, but I think it's around $600000 for a 40TB one. That's what you
do, if your data is really worth keeping.

Could a linux kernel be modified to make it's SCSI hostadapter respond to
SCSI requests for multiple IDs, and generally act like a bunch of disks, as
this thing can ? That would rock...

Kate

-- 
Microsoft - the best reason in the world to drink beer

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/