On Fri, Jul 14, 2000 at 02:29:10PM -0700, Greywolf wrote:
> In light of this, isn't this along the lines of what a Beowulf cluster
> does?  And if that's so, there might be some information available from
> the Beowulf team/symposium/whatever.

Um, I suppose, in the larger sense. The Beowulf, though, paper has
nothing about using one machine's RAM as a backing store for another
machine, which is the specific topic I'm working on right now (though
it's part of a larger clustering effort). Do you know of their doing
something similar?

There are a few problems in relying to heavily on anything out of
Beowulf:

1) Beowulf is Linux-specific.

2) Beowulf is i386-specific.

3) Beowulf presumes each node has identical hardware, and load
balances blindly based on that assumption.

I want to run on NetBSD, I want to run on every architecture supported
by NetBSD, and I want to place processes intelligently based on the
ability of the nodes. (... and I want to distribute control data,
backing store, and disk over all the nodes without any central master
so that the whole thing is tolerant of a node failing or a new one
being added at random)

Beowulf does a very good job at what it's designed for, but that's
only similar to what I'm designing in that it's a parallelized
cluster. There are a *lot* of details that are different.

       ~ g r @ eclipsed.net