redundancy of SCSI controllers possible?

Tue Mar 23 09:41:01 PST 1999

On Tue, 23 Mar 1999, Carlos Carvalho wrote:
> Torbjorn Lindgren (tl at fairplay.no) wrote on 23 March 1999 13:10:
>  >Single point of failure to take the box down, yes. The data is still
>  >there, so it's different from a multi-disk failure.
> 
> I'm talking about writing, not reading. If you have two disks on each
> controller, if one controller fails those disks won't be updated, but
> the others will. So the info is only partially written, and I'm not
> sure the data is there. Or is the raid code smart enough to detect
> this? If so, controller failure isn't a problem at all.

I would hope that "serious" RAID solutions to have some kind of checking
mechanism to use to handle this. Perferably on-line so you can run the
check/verification while the system is running.

>  >You have to decide what you aim is, and how much you are prepared to pay
>  >for it. What if the PCI bus fries? What if one CPU fries? Etc etc...
> 
> These kinds of failure don't break the synchronization of the raid.

Actually, I would expect that they could do that, the time window is
just much smaller, unless someone is doing really aggressive cacheing
(which is more or less require to get any performance...), but I can't
see how to avoid it altogether, unless you do some kind of transaction
based RAID solution (so you can roll back). A hardware system with
battery backup should also be able to handle it easily though.

I would be surprised if most software RAID used that kind of transaction
logging required to handle this.

>  >IF the aim is to make sure that the box can stay up with single controller
>  >failure I would make sure I have enough controllers to never put more than
>  >one disk per RAID set on each controller. 
> 
> This isn't possible, there aren't enough PCI slots.

Get more PCI slots :-) Supermicro[2] has cards with up to 9 PCI slots
and two U2W SCSI channels on the motherboard! And it has an AGP slot
to put the graphic card into...

I'm not sure if it works with Linux though, or if it requires I2O to
work? Looks like Alan working on at least some I2O support, so perhaps
all isn't lost even if it does require I2O. I suspect that it works
fine even without I2O support, but I don't really know.

You ought to get MOST of the benefits if you use fewer multi-channel
cards instead, Initio[1] for example has dual and quad channel SCSI
cards! They also has a cute dual SCSI channels plus one fast ethernet
card... (only for UW thought, they only have a single channel U2W
card).

Adaptec also has multi-channel products, up to two? channels,
including a dual channel U2W controller. Mylex seems to boost both
conventional multi-channel cards, and multi-channel RAID cards (up to
three U2W channels).

Other RAID controller vendors also has multi-channel cards, and I
think at least DPT does support multi-card RAID's (the Millenium
boosts up to 3 U2W channels if I read it right).

These are the ones I could think of at the moment, there are
undoubtably others.

You don't get protection from all controller failures, but it should
protect against trouble with one SCSI chain, which IMNSHO is much more
likely. As usual it's a question of money...

>  >For large configurations this might be doable without large costs, by
>  >using multiple RAID sets (multiple disk on each controller, each one from
>  >a different RAID set).
> 
> You want even more disks than Doug :-)

I don't have any plans on building any such behemots here, but it's
fairly easy to figure out what would be needed.

If I still worked at my previous employer I would probably be looking
for something much larger than what I support now, either multiple
50+++ GB systems (there are possible internal sub-divisons that could
be used) or one much larger systems (300-700 GB perhaps?) given what a
much smaller number of people manages to consume here (and the way
everone used space there)

With large systems it's easier to do this kind of things, using 5 or
more SCSI channels isn't a big problem there :-) (Say ~576GB usefull
space as 18GB*5*8 or 36GB*5*4, both with 5 channels).

>  >This should work as long as the failing controller doesn't do something
>  >really bad (like blocking the PCI bus until it's removed!)...
> 
> Again, this doesn't break the raid.

I suspect that it could do that with the right timing. Admittedly only
for a single write, unless someone is doing aggressive cacheing, but
on the other hand the performance usually sucks unless you DO use
aggressive cacheing with higher RAID levels...

For really sensitive data external hardware boxes usually are the way
to go IMNSHO, with redundant links to them. Be prepared to pay a LOT
though.

1. http://www.initio.com/
2. http://www.supermicro.com/

-- 
Torbjörn Lindgren
Network Manager, FairPlay International AS
E-mail: tl at fairplay.no

To Unsubscribe: send mail to majordomo at FreeBSD.org
with "unsubscribe aic7xxx" in the body of the message