Re: sharing SCSI disks
Keith Rohrer (kwrohrer@uiuc.edu)
Thu, 20 Mar 1997 19:34:44 +0000 (GMT)
> On Wed, 19 Mar 1997, Erik Walthinsen wrote:
> > On Wed, 19 Mar 1997, Eric Hoeltzel wrote:
> > > This just boggles my mind. If I'm considering the inferences
> > > correctly, for short distances if you need far faster than 100
> > > megabit netcards between a couple of boxes then you can use a
> > > few scsi host adapters? Perhaps a couple in each box and load
> > > balance them with the eql driver?
> > It'd make an impressive clusters connection. A couple of 20MB/sec rings
> > doing IP... Fast and cheap. Even at $200/card for good ones, you could
> > outfit a two-system cluster with 40MB/sec max theoretical bandwidth for
> > about $900 (including cables). Try doing that with ether! MyriNet
> > (1.28Mbps each way) can't even compete, especially at $1.400/card.
Actually, bang-for-buck I was thinking more along the lines of $50
NCR/Symbios 53c810-based cards (Fast, not wide, not ultra, but PCI).
This would compare well to cheap fast ethernet cards, even better if
the systems were already SCSI and shared one ultra or ultra/wide (tho
wide cabling gets expensive) bus. You wouldn't get the cable distance
(and the cables would be more expensive), but it would likely beat out
both fast ethernet cards with a $1000 fast ethernet hub and a 2-card-
per-machine point-to-point forwarding ring setup (assuming your fast
ethernet cards use the wires in a way which enables crossover cabling).
> > Of course, you have to assume the bus doesn't go nuts having two systems
> > on it, though someone posted a few hours ago that they have two machines
> > safely sharing devices in both enforced and voluntary read-only modes.
> > There's also quite a bit of overhead involved, and you're limited to
> > machines sitting side-by-side or back-to-back, but still...
> Actually, I wasn't even thinking of having disks on the scsi bus at
> all. Just a host adapter (or more) in each box instead of ethercards.
All it would limit would be the distance between boxes (not that long
SCSI cables are cheap anyway); as long as the SCSI ID's don't collide
and the SCSI bus bandwidth is excessive, it shouldn't even be a big
performance hit.
> > Anyone with hardware hanging around they could try this out with?
One of these days...
> Unfortunate that Linux is typically installed on lesser boxes because
> it runs so much faster. All my good hardware goes to compensating
> for NT's sluggishness. The concept sounds fascinating however. Could
> the eql balancer pull this off at these speeds?
I've heard someone say that eql wasn't likely fast enough even for
ethernet. I don't have any evidence to back it up, though...but it
was designed to balance (slow) serial connections.
Keith