parallel processing smalltalk

Eliot & Linda elcm at pacbell.net
Wed Dec 9 19:09:33 UTC 1998


Aik-Siong Koh wrote:
> 
> I use smalltalk in an engineering software for motion simulation. Smalltalk
> has been excellent for my programming productivity, but now I am looking
> for computational speed. Are there examples of smalltalk taking advantage
> of massively parallel processing? I would also be interested in any
> smalltalk plus C, C++, or FORTRAN combination which takes advantage of
> parallel processing.
> 
> Any suggestion on how to maintain the programming productivity of smalltalk
> and yet achieve the computational speeds expected of engineering/scientific
> software is greatly appreciated.


I know of work in the fiunancial modelling area which uses clusters of
PCs using VisualWorks and sockets to do floating-point intensive
computations for "advanced" financial instruments.  The cluster was
interesting in that some of the workstations weree in Now York while
others were in London, and when the load in London reduced as people
went home calculations started in New York would spread to London!  Note
that the application doesn't use any external code; all calculations are
in Smalltalk, and current Smalltalk implementations have very poor
floating-point performance becase of the need to box floats, requiring
costly instantiations for each result.

I'm very impressed by the low-cost supercomputer-class clusters that
Laurence Livermore are building, i.e. Beowolf with Intel PCs and Avalon
with Digital AXP machines.  These clusters spend more on each ethernet
card than on the rest of each node.  They use gigabit ethernets and
implement distributed shared memory, and achieve high performance at an
order-of-magnitude less cost than the fastest SMP machines available
from Sun and SGI.  The same architecture could be used to host similar
models to that used by the financial modelling application above. 
Smalltalk's object model has shown the way in the distributed oo area;
one can trace CORBA back to distributed Smalltalk work done in the early
eighties.  In the commercial sphere DST and SmalltalkBroker already
provide suitable starting-points for massively-parallel Smalltalk
architectures.
_______________,,,^..^,,,_______________
Eliot Miranda, ParcPlace





More information about the Squeak-dev mailing list