[ILUG] don't ya just hate it when...

[Kenn Humborg]

> > So in effect with this scheme a large hardware coherent NUMA machine
> > would essentially be X seperate linux kernels (X=# of nodes) running a
> > distributed smpfs to coordinate access to memory, IO, inodes, etc..
> 
> I don't see how that will make your kernel/kernels any more 
> efficient, you're
> basically applying a distributed computing model to parallel 
> computing which
> involves more work
> 
> > Now imagine one day someone actually implements such a thing for
> > linux. The only difference between such a machine and a bunch of PCs
> > would be the hardware memory access...
> > 
> > just one more step of someone writing code to implement the NUMA stuff
> > over, eg ethernet or UDP. And hey presto... distributed linux will be
> > born.

I've missed some of this discussion (emails got downloaded to
my home PC before I had a chance to read them from work), but
what you're describing seems quite similar to VMS clustering.
A VMS cluster has a cluster-wide disk/tape namespace, a 
cluster-wide process ID namespace, a cluster-wide distributed
lock manager namespace/database and cluster-wide logical names
(think of environment variables on steroids).  Oh yeah - I forgot
cluster-wide batch and print queues.

If you added a generic intra-cluster mechanism for remoting
_any_ I/O devices (which, to a first approximation sounds
pretty straightforward), such as terminals and network 
interfaces, then you've basically got cluster-wide access
to _everything_ except local memory on each node.  And that
could probably be hacked through the page fault handlers...

Hmmm...  Must get a cluster up and running at home some day
and play around with it :-)

Of course, to get Linux to this stage, we're going to need
industrial-strength distributed file systems.  GFS looks nice...
(www.globalfilesystem.org) 

Later,
Kenn