pack@acd.acd.ucar.edu (Dan Packman) (12/30/90)
In comparing a given application (for simplicity's sake, running on an
unloaded system) on many platforms, it might be nice to have a set of
parameters (ideally empirically determined) that would allow its comparison
to standard benchmarks with parameters determined for them as well. The
numbers that come to mind (any suggestions?):
1. How io intensive? Something like math ops/megabyte. Perhaps
another number indicating numbers of io per megabyte. Presumably
this number would indicate that all ten SPEC benchmarks are
not very io intensive (eg. a low number).
2. How floating point intensive? Something like integer ops/floating ops.
3. How vectorizeable? A percent figure might be nice as well as a typical
vector size.
4. Scale of problem. We need some measure to set machines with large
and efficient data caches apart from others.
If we had say six to twelve numbers to characterize programs, then we
could find a published benchmark closest to the application of interest.
We could accurately predict our application's performance on reported
platforms. Similarly, if we had many benchmark results for a given
platform covering a wide range of these characteristic numbers, we could
predict the performance of a wide variety of applications.
Any ideas on how to devise these characteristic numbers and how to
measure (or calculate) them for a given program?
Dan Packman NCAR INTERNET: pack@ncar.UCAR.EDU
(303) 497-1427 P.O. Box 3000 CSNET: pack@ncar.CSNET
Boulder, CO 80307 DECNET SPAN: 9.367::PACK