[ale] mosix clusters?

[Jeff Hubbs]

On Sun, 2002-06-23 at 17:51, Christopher Fowler wrote:
> I would have the master use disks and the slaves share via NFS.  Besides
> does'nt Mosix require the use of an API to be taken advantage of?
> 
> Chris
> 
> On Sun, 2002-06-23 at 17:39, Stephen Turner wrote:
> > do mosix cluster nodes share hdd space? as one hd or do they mirror it so
> > its backed up?
> > 

I can actually take these questions, kinda... :-)

Mosix sharing HDD space: Mosix has what it calls the Mosix File System
which makes it possible for every node to see every other node's disk
space.  I haven't fired this up yet so I'm not entirely coversant with
its operation...YET (I am currently working on a Gentoo/OpenMosix "Node
0" right now).  

Having the nodes use NFS is do-able but MFS makes it unnecessary, or so
I'm starting to understand.  Exception:  you want your entire Mosix
cluster to see filespace elsewhere, like on a file server, if no other
reason than so you can get away with smaller hdds on your nodes.

Need for API:  Chris, you're thinking of Beowulf-class clustering. 
Mosix' big draw is that it requires little attention in the way of how
you code or what you run, meaning that you can leverage Mosix for COTS
software or stuff you write yourself with little or no special coding
considerations.

What I would like to be able to use Mosix for someday is a business
computing situation where the computing operations are generally
batch-like, i.e., multiple independent jobs that would ordinarily
execute serially on a single machine.  All such jobs might operate upon
the same data files or RDBMS.  Under VMS' batch mechanism, you'd
establish one or more queues and configure them to hold X jobs and
execute Y at a time; your choice of Y tended to determine how hard you
ran your machine, which you of course would be balancing off against
completion time.  If your batch jobs build up faster than the machine
can finish them, you're not in very good shape.  

To me, the great thing about trying to handle a computing problem like
this with Mosix is that you can implement your batch mechanism (or
equiv.) on Node 0 and you can add or take away more horsepower as you
need it or don't need it.  Wait, it gets better.  You can take a pile of
old computers that some outfit lays out by their dumpster and throw them
onto your cluster like gerbils on a wheel.  And before you go, hey, what
good are ten free Pentium 90s when I can get a 900MHz Duron cheap,
consider that the ten free Pentium 90s have a 320-bit-wide memory pipe. 
Now, if you've got money to spend, then, yes, you can buy some number of
really fast and compact machines, but you also can amass a lot of power
just with other people's junk! 

Of course, this could start to fall apart if your application is so
network-traffic-heavy that the cluster doesn't scale well (ISA-bus
Gigabit Ethernet cards are hard to find :-)) .

- Jeff








---
This message has been sent through the ALE general discussion list.
See http://www.ale.org/mailing-lists.shtml for more info. Problems should be 
sent to listmaster at ale dot org.