roberts@NBS-VMS.arpa (ROBERTS, JOHN) (08/25/86)
I would seem that a new problem has appeared in recent years in the
collection of comparison benchmarks, in relation to reconfigurable
multiprocessor systems. Several of the systems mentioned can be run
with at least a 10:1 range in number of processors, with additional
options on memory size, etc. I assume the numbers posted were "per
processor", though this is not explicitly stated. Also, multiprocessor
systems do not generally get full linear speedup in normal applications,
and the amount of memory available for a task can greatly affect
rate of execution.
It would be helpful if the multiprocessor test results told more about
the specific configuration under which the test was run. For measuring
effective speedup, etc., it would be nice to get results for a minimal
configuration and for a larger or maximal configuration. Of course,
such measurements would not fully describe the characteristics of these
systems, but they would be more comparable to the measurements of the
other systems.
(I do not know the extent to which the Dhrystone benchmark can be
parallelized, but running multiple copies simultaneously may be useful.)
<Standard disclaimers.>
John Roberts
roberts@nbs-vms.ARPA
------