[ale] Sportin' SCSI/RAID

[Jeff Hubbs]

Eric -


When a VMS cluster boots, all the nodes execute the same boot procedure
(although there can be node-specific procedures, there usually isn't very
much in them) and, best as I recall, each machine does execute MOUNT/SYSTEM
commands against the same volumes.  

I'm not really trying to bemoan or belittle anything, but I *would* like to
see Linux achieve parity with VMS in the shared-disk cluster area and with
NT in the general clustering area as well.  I consider Beowulf-class Linux
clusters to be kind of a "special case" of clustering.

Both of the VMS cluster environments I've worked in had the goals of 1)
providing some measure of fault tolerance (i.e., one machine dies and you're
not totally hosed) and 2) expanding capacity for interactive users.  Neither
achieved full failover in the way that NT and some UNIX cluster solutions
hope to do and neither used multiple nodes to carry out some divisible
computational process in parallel, so, by some measure, VMS clustering was
left wanting too.

My particular Linux clustering idea really wasn't really all that concerned
with clustering per se because I expected the two machines to be
asymmetrical in both power and function.  The only real "cluster-like
feature" I was trying to introduce was the shared SCSI channel.  I think I
could accomplish what I was trying to do through NFS, but I wanted to avoid
relying on network stuff for disk access.

BTW, in an NFS export/mount situation, does the mounting machine do any read
caching, or does all that take place in the exporting machine?

- Jeff

> -----Original Message-----
> From: Eric Z. Ayers [mailto:eric.ayers at mindspring.com]
> Sent: Monday, December 27, 1999 8:09 AM
> To: Jeff Hubbs
> Cc: Eric.Ayers at mindspring.com; Jeff Hubbs; ale at ale.org
> Subject: Re: [ale] Sportin' SCSI/RAID
> 
> 
> 
> If I had a dime for every time I heard the guys familiar with VMS at
> our office bemoan the shortcomings of UNIX clustering!
> 
> It is likely that the SCSI reservation feature is optional...
> 
> One difference is, that in VMS there is not a large filesystem cache
> like there is in UNIX, which would make it almost impossible to keep
> the filesystem consistent if mounted on multiple machines.  We have
> installaons with many hundreds of megabytes on outboard RAID
> controllers.  Compaq is introducing a new version of their clustering
> with Tru64Unix (was Digital Unix) version 5.0A. But my understanding
> is that the way they provide access is through a cluster interconnect
> (they call it memory channel) - still only one machine actually mounts
> the drive. 
> 
> -Eric.
> 
> Jeff Hubbs writes:
>  > Eric -
>  > 
>  > My only real clustering experience is under VMS, in which 
> case the systems do all
>  > mount the same volumes and in some cases do it over a SCSI 
> channel.  The DEC
>  > AlphaServers I worked with most recently had outboard RAID 
> controllers and shared a
>  > set of disks between them, cut up into several RAID 
> arrays.  I'm not trying to
>  > contradict you, though - I have every expectation that 
> there was some mechanism that
>  > made one controller or the other "lay out" while one 
> worked the drives.  At the time
>  > I got laid off, we were working on an Ultra-SCSI-2 SAN 
> connecting three nodes.  But,
>  > yes, it's looking like in Linux at least, what I'm
>  > finding is that there's not a similar mechanism, at least not yet.
>  > 
>  > Actually, I was not looking at this in terms of a cluster 
> per se - I was trying to
>  > work out a possible arrangement of machines for a 
> public/private Internet Web server
>  > in which one machine ran Apache and MySQL (but only 
> reading a database) and the other
>  > ran Apache and MySQL (read/write).
>  > 
>  > - jeff
>  > 
>  > 
>  > "Eric Z. Ayers" wrote:
>  > 
>  > > Hello Jeff,
>  > >
>  > > Most UNIX clustering solutions don't allow both machines 
> to mount the
>  > > same SCSI target,LUN at the same time.  In fact, the 
> SCSI protocol has
>  > > this concept called a reservation which explicitly 
> prohibits two host
>  > > adapters from talking to the same device simultaneously.
>  > >
>  > > -Eric.
>  > >
>  > > Jeff Hubbs writes:
>  > >  > Does anyone here have any experience with sharing a 
> RAID array between two
>  > >  > Linux systems over a common SCSI bus?  What is 
> required?  What are the
>  > >  > gotchas?  Is it less of a problem if one of the two 
> machines only reads from
>  > >  > the RAID volume?
>  > >  >
>  > >  > If the shared RAID array in question held, say, a 
> MySQL database, how would
>  > >  > both machines access it (again, presuming one machine 
> only reads)?  An
>  > >  > instance of MySQL on each system?
>  > >  >
>  > >  > - Jeff
>  > >  >
>  > >  > ** Happy Holidays to the ALE! **
>  > >  > --
>  > >  > To unsubscribe: mail majordomo at ale.org with 
> "unsubscribe ale" in message body.
>  > > --
>  > > To unsubscribe: mail majordomo at ale.org with "unsubscribe 
> ale" in message body.
> --
> To unsubscribe: mail majordomo at ale.org with "unsubscribe ale" 
> in message body.
> 
--
To unsubscribe: mail majordomo at ale.org with "unsubscribe ale" in message body.