[9fans] Re: Loosely-coupled multi-processor CPU servers?

[9fans] Re: Loosely-coupled multi-processor CPU servers?

[dmr]

Fritz, then Mirtchovski ask

> This discussion brought me to a problem I've been looking at for a
> while: I've been studying the problem of building a vest-pocket
> rendering farm.  For my purposes, I'd like a board or array of small
> boards with perhaps 16 very fast, cost-effective processors, each with
> perhaps 128 or 256MB of RAM, and a fast interconnect--30MB/sec or more.
> Is this--or module-level parts to build it--an off-the shelf item?

 > 1. How easy would be to add new cpu servers to a plan9 network? Can they
 > be 'loosely coupled' in the sence that when need arises new machines will
 > be added to an existing environment and removed later? (i have heard of
 > something like that in pixar's rendering farm)
 > 2. (Pardon if this already exists in documentation -- I couldn't find
 > it...) How easy it would be to migrate processes from one cpu machine to
 > another, if at all possible?

You can certainly have plenty of CPU servers; we have two
main ones and others that can be coopted readily.

What isn't in the distribution is an automatic load-balancing
system (though it would be a SMOP for the  cpu  command
to choose one based on some criterion).

What is definitely not soon in the cards is process migration
(transporting a whole running state).  Namespace, yes.

However, something similar to the "rendering farm" model
has been built by Gerard Holzmann (with McKie et al.)
in his FeaVer system.  The name has to do with Feature
Verification, specifically for almost-automatic formal
model-checking for telephone and other systems: see
the first two items in

http://www.cs.bell-labs.com/who/gerard/abs.html

The BLTJ paper on FeaVer and the 15 or 16 CPU rack
that was built for this is about to appear.

	Dennis