[comp.lang.prolog] DelPhi Prolog Machine Progress

wfc@cl.cam.ac.uk (William Clocksin on jenny) (01/04/89)

Readers of this bulletin board may be interested to know that a
20-processor DelPhi machine has been up and running for a month
or so.  For the usual nondeterministic benchmarks such as
N-Queens, the speedup curve is linear up to the amount
of parallelism in the problem, and then *completely flat* up to
the number of processors.  This is the behaviour people have
been looking for, and it suggests that the DelPhi design is
not prey to the bandwidth contention that is characteristic of
the various shared-memory designs people have been discussing.
As expected, the overheads of the DelPhi principle
add only about 30 percent to a single-processor version, and this
overhead is completely recouped (and then some) when a second
processor is used.
Our small benchmarks run out of parallelism at about 10 processors,
but we have some electrical design problems to try, which should
use up all the processors.