Re: Wolfram Research: Personal supercomputing is here.

> Looks at though your $4,085 will buy you a 150% increase in speed (see
> http://www.wolfram.com/products/personalgrid/advantages.html ).

> Now let's think - slightly faster CAS or 3 Caribbean cruises? Hmmmm...


Apple have just started shipping the G5 Quad-core (two dual-core
processors), so I read this release with considerable interest.

On a 4 processor system:

* Wolfram devote 1 copy of Mma running as a 'master' running
on processor number 1,

and then:

* 3 copies of Mma, each running on a separate processor core,
and each acting as slaves doing the grunt work that
is handed out to them by the master.

I imagine this is a reasonably easy way to convert an app that has been
written for single-core processing to reap the advantage of
multi-processor machines. And, as one would imagine, Wolfram state that
it works best with problems that can be broken up into smaller bits. And
even though Integrate (apparently) can't be broken up (odd - one would
think that integration of Piecewise funcs would be an ideal contender),
I often tackle integration problems that require solving 12 or 14
separate integrals which could presumably be split up veddy nicely
across 3 processors.

Despite my considerable salivation, I somewhat doubt that this
particular implementation will be a viable long-term solution: First,
to take advantage of it, you need to write special code: so, instead of
Mapping a func across some data, you need to call a special func called
ParallelMap (or some such name) to tell the master kernel to split it up
across the slaves. And that means, if you are developing code for other
users, your code is no longer machine-independent. Second, as dual- and
multi-core computers become increasingly mainstream over the next couple
of years, I think it is clear that end-users will expect their software
to automatically take advantage of the grunt power they have purchased,
and not want to purchase multiple copies of their software to run on
each processor, anymore than someone buying a single processor 4 Ghz
machine will not want to pay any more for their software than someone
who buys a 2 Ghz machine. Users will just want it to 'work'.

Neat - but early days :)
.

Quantcast