lnz at dandelion
Dec 9, 1998, 10:55 AM
Post #5 of 18
Date: Wed, 9 Dec 1998 05:17:15 -0800 (PST)
Fw: Help! Problems with IP address takeover on 2.0.36
[In reply to]
From: David Lang <dlang [at] diginsite>
I have now seen several people make the comment that SCSI does not allow
more then 2 machines to connect to it. where is this limit? As I
understand SCSI there is nothing that would prevent you from using wide
scsi to hook up 15 computers to one drive and, assuming they were all
accessing different partitions, they should even be able to access the
I do not believe there is any specific limit on the number of initiators. On
4-way Sequent clusters, there most definitely were 4 initiators and 12 target
devices on their differential SCSI buses.
From Agent Tatum <josh [at] ai> Wed Dec 9 19:57:19 1998 
From: Agent Tatum <josh [at] ai> (Agent Tatum)
Date: Wed, 9 Dec 1998 13:57:19 -0600 (CST)
Subject: SCSI questions.
In-Reply-To: <366E902E.E450EA22 [at] bell-labs>
Message-ID: <Pine.LNX.3.96.981209131730.3476A-100000 [at] ai>
On Wed, 9 Dec 1998 alanr [at] bell-labs wrote:
> David Lang wrote:
> > -----BEGIN PGP SIGNED MESSAGE-----
> > I have now seen several people make the comment that SCSI does not allow
> > more then 2 machines to connect to it. where is this limit? As I
> > understand SCSI there is nothing that would prevent you from using wide
> > scsi to hook up 15 computers to one drive and, assuming they were all
> > accessing different partitions, they should even be able to access the
> > drive simultaniously.
> > David Lang
> You're right, at some level. The problem is cable runs, cable length,
> etc. are usually problematic in practice. Fast SCSI has relatively
> short cable length restrictions (though I can't recall what they are).
> By the time you make neat cable runs, it's a problem. Also, SCSI is
> difficult or impossible to hot-plug, so that if one machine comes down
> to be replaced or significantly messed with, it may require the SCSI
> activity on all the other machines to be stopped at least momentarily.
> The number two probably dates back to 8-address SCSI, but is probably
> pretty close to the practical constraint in faster SCSI, because the
> cables have to be shorter, and things are more tricky.
> -- Alan Robertson
> alanr [at] bell-labs
I have personally set up 8-node single bus HACMP clusters attached to
various RAIDS (EMC, CLARiiON, etc.). AIX has a concurrent mode that, with
the use of a lock manager, allows as many nodes as you can fit on the bus
to write to a particular raw logical volume. The most frequent problem
that I've run into has been resets on the SCSI bus and how well the RAID
responds to them. When working properly and using Y-cables on the backs of
all of the nodes, I/O shouldnt stop but will quite possibly pause
momentarily (5-10 seconds) while the rejoining nodes initiates the
adapter/disks. Most RAIDs now days have several SCSI ports, all of which
can be configured to see all of the disks which. This allows for you to
even yank out and replace a bad cable without disturbing I/O on the other
nodes. Bringing down a system for maintence manually should not cause I/O
to stop unless you are also switching an application from the node being
Unexamined faith is ignorance.