Subject: Re: multi CPU uvaxII
To: Brian D Chase <bdc@world.std.com>
From: A. Sporner <asporner@eagle.ibc.edu>
List: port-vax
Date: 03/17/1998 14:19:49
Hello,

> On a slightly related note, I think I've heard it discussed somewhere at
> some point about chaining VAXstations together over their SCSI busses. It
> may have been with one of the VAX/Linux people.  In the case of the
> VS3100's with dual SCSI controllers would it be possible to say take two
> machines and patch their SCSI/B (2nd) SCSI controllers together?  Then set
> up each controller with a unique SCSI ID on the shared bus.  Then you
> could build a driver for the SCSI devices to make them look like
> point-to-point network interfaces.  Basically use them for passing data
> between VAXen at "high speed".  On top of this you could use something
> like PVM or MPI to run parallel computations across the machines.  And
> then lets say you take the dual SCSI controller machines, you chain then
> together in a big long bus topology, or even a big ring :-) 

IFAIK, there is a general problem with LINUX (off topic, sorry) in that
there has been a problem with unsolicited SCSI requests. At one point and
might also be current, if an unsolicited request arrive, the kernel would
flat out panic.  The assumption being, I didn't ask for this.  I know that
work is being done at that project to completely overhaul the SCSI subsystem
to overcome this.  

As a matter of practice, aside from VMS. Unix machines I have worked on
have used the Shared SCSI busses now for about the last four years.  The
difference is that they use differential versus single-ended.  Our production
machines all have at least 10 shared SCSI channels.  The concept is very
good!

> 
> So, is this a technically feasible abuse of SCSI?  And if so, how hard
> would it be to make driver for the SCSI devices that made them look like
> network interfaces?  AND if we could intersperse a SCSI disk drive or two
> in the chain, we might have a faster alternative to ethernet for shared
> disk storage.
> 

The mixing of disk and TCP/IP protocols is already being done on fibre
channel (http://www.emulex.com/press/lp7000.html).  It would be a cool
think to do it with SCSI.   

Regards


Andy