Cluster update

[david nicol]

On Tue, 2002-12-03 at 11:40, Jared wrote:

> I don't know anything bout Mosix
> and Beuwolf except that Beuwolf needs identical boxes
> and Mosix can use a mix of different configurations.
> I know others on the list know much more!

I built two MOSIX clusters for UMKC.

A MOSIX cluster fits the loose definition of a "beowulf"
as it is a cluster running Linux.

What MOSIX gives you is: your forking multithreaded process,
(such as a parallel make -- MOSIX is really good at large builds)
gets distributed around the cluster transparently.

What it doesn't do is forward IO very well.  Which is still fine
if you're compiling: read all the input, chew on it a while,
write some output.  As long as your processes match that paradigm
MOSIX is good.

What I would like to do with the cluster includes:

	1 sell the machines in it to end users who come into the
	shop.

	2 have it as the core of a wider cluster that can
	 be joined by anyone with broadband who is so inclined and
	 signs up

Gerald's suggestion of doing build QA for sourceforge etc. works
well with #2.  Parallel builds often break where non-parallel builds
don't break, since a lot of developers don't have dual-CPU machines
so they never bother to test their builds with -j5 or -j10 or even
with -j2 which are the make switches for building using 5, 10 and 2
concurrent compilations, respectively.

While I was working with MOSIX at UMKC a couple of nagging problems with
it got onto my long, long eventual todo list.  They are listed in my
postings to the linux-cluster mailing list in fall 2000 and do not
belong in this message.It is conceivable that grants to achieve those
enhancements to cluster technology might be obtainable.  Then we could
hire ourselves to implement them.

There's another project I mentioned recently here on kclug list, the
universal e-mail sender whitelist, that nobody really responded to. That
project would require bandwidth rather than CPU.  Although I'm not sure
how much: DNS is pretty compact, and scales well.  Does anyone know
how much bandwidth is used by the main RBL DNS?   Furthermore if we
(the whitelist provider -- pay2send.com) charge for smtp server
membership to use the whitelist heavily, there's an income source for
the cluster project.

-- 
David Nicol, independent consultant and contractor            312 587 2868
                       Howard Zinn quotes Thomas Paine.  Who do you quote?