Darren Hobbs wrote:

>Hi,
>
>Loath as I am to continue in the multiprocess / multithreading vein,
>I'm going to anyway :-)
>
>We're starting to see machines like these appear in the market:
>http://www.orionmulti.com/products/
>
>  
>
A quick look at their add shows that these processors are loosely
coupled (high speed network connection between them), not SMP.  So, the
quickest option that I can think of is writing an MPI or PVM adapter for
Squeak.  PVM is a relatively small protocol so it might be easier to
use.  I don't think that Dave's suggestion of forking processes will
work since I don't think that these 96 processors are scheduled by a
single kernel.  Instead you need a mechanism to start processes on the
"remote" processors, communicate with them and shut them down when
you're done.  These are the things PVM/MPI/LAM/MPICH are designed to
do.  Alternatively you can use one of the squeak object marshaling
protocols to communicate and write your own tools to manage the remote
processes.  Since PVM and MPI are "strongly typed" you might find this
solution more palatable.  As for memory, these machines have one RAM
DIMM per node (oddly enough it looks like 48Mb DIMMs) and I'm sure that
the kernel and support processes on the worker-bee nodes are very lean
so 30-35Mb image would certainly be manageable.  You could also look
into the cost of upgrading these to 128Mb DIMMs.  As one final comment,
these machines could easily be used in an apache load-balanced
environment.  One image per node w/ the standard apache
mod_proxy/RewriteEngine hacks to get it to balance the load between
them.  No inter-image communication needed.  This would make a great
Squeak+Seaside server right out of the box!

David

-- 
C. David Shaffer
http://www.cs.westminster.edu/~shaffer
http://www.shaffer-consulting.com