[ILUG] Have you had your daily dose of reality?

[Kenn Humborg]

> larry mcvoy had a post on l-k a while back. according to him single
> image parallel hardware systems (like the e10k> are bad. he reckons a
> large cluster of single image systems that can share resources are the
> way to go, rather than finely locked complicated smp kernels.
> 
> one idea he put forth was an "SMPfs" a sort of nfs for kernels in a
> cluster to claim locks on resources in other machines. Sounds rather
> cool.. eg kernel on host1 could claim memory on host2 via smpfs. that
> kind of thing. So benefits of single system for apps's, without most of
> the complexity of a single system kernel.

That sounds very much like a distributed lock manager (hmmm,
now what OS used that...  oh yeah!  VMS - that's it :-)

The lock manager in VMS was a seriously sophisticated beast.
You basically had a hierarchical lock namespace that was
cluster-wide.  Once you had a 'handle' to a lock, you could
request one of 6 levels of access to it (none, various 
combinations of allow/deny read/write and exclusive) and
change your level at any time.  You could block on the
level change, or tell the lock manager to let you know when
your request is granted.  You could also ask to be notified
when another process requested a level that was prevented by
you (a neat form of intra-cluster IPC, because locks could have
a 16-byte value block for use by apps).  It also provided
automatic deadlock detection and resolution.

The really neat thing was the distributed aspect.  Locks would
be mastered on whatever node was making most use of the lock
and migrate to other nodes as usage changed, also taking into
account the current load on the node.  If multiple nodes we using
a given lock, the non-master nodes would have slave copies of
the lock.  In the event that a node died, any other nodes that 
were using locks mastered by it negotiate a new master.

The main use of these locks was in the file system and in the
record management system to ensure consistency on disks that
could be accessed by multiple nodes.

This is the direction that the Linux Cluster Cabal are heading.
That's going to be _cool_.

Later,
Kenn