Re: Clustering for Linux 2.3.x?!

[slurn@griffin.engr.sgi.com (Scott Lurndal)]

In article <7bk2gs$acocm@fido.engr.sgi.com>, you write:
|> I like the idea of clustering, but why not go a step forward.
|> 
|> Build a virtual machine which is spawned of many systems. Connect the kernel
|> which a fast network ( giga ethernet ?? scsi ??? firewire ?? ).
|> Greate would be featers like
|> 
|> add an other machine to the cluster "online".
|> remove a machine "online"
|> move "running" applikations from on machine to an other
|> run a "mirror" application, which is an application which runs on two "real"
|> machines.

Would you be extremely surprised to find that such a machine
has, indeed, been built?

The Unisys OPUS system was just such a system.  Consisted of
4 to 128 nodes - each node with a Pentium Pro, 512M-1G ram, 
2 fast/wide scsi channels, 2 100baseT enets and connected
with the Intel MESH (a la Paragon) backplane.    The operating
system (SVR4/MK: derived from Chorus and parts of SVR4.2MP) provided
a true single system image in all respects across all nodes
that made up the system/cluster.  

The OS was just as happy running on a set of PC's connected
with fast ethernet (albeit much slower than the intel Mesh
backplane). 

Process migration between any of the nodes was fully supported
at any time.   All unix semantics were preserved, including 
coherency of files when opened by processes running on different
nodes (even when sharing the off_t value), as well as read/write
atomicity guarantees and a global /proc implementation.

If a node went down, raw disks would automatically fail-over
to another node (using dual-ported disks).

Unfortunately, the system tended to compete too well
with the proprietary mainframes and was canceled after
20 or so shipped to customers.  C'est la vie, 10 years
of work down the tubes.   I wish that it were open source....

It would work today very well on clusters of small ia32 boxes
with a fast interconnect such as gigabit ethernet.

|>   if one computers fails a takover without any loose of any data and computing
|> time is possible.
|> 
|> this would be a real transparent cluster.

Yep,   and it sure runs oracle OPS really well for decision support.

scott lurndal
silicon graphics, inc. 		(I speak for myself)
	
-
Linux-future: thinking about the future of the Linux kernel
Archive:      http://humbolt.nl.linux.org/lists/
Wish list:    http://users.ox.ac.uk/~mert0236/linux-future.html