[HPGMG Forum] Performance Versatility slides from Monday
mfadams at lbl.gov
Sun Nov 30 18:12:14 UTC 2014
On Sun, Nov 30, 2014 at 12:27 AM, Jed Brown <jed at jedbrown.org> wrote:
> Brian Van Straalen <bvstraalen at lbl.gov> writes:
> > I don’t think it is this groups charter
Yikes, we don't have a charter do we :)
No really, our charter should be something like: 1) make a good top500
ranking metric, which means distinguish networks to start, and 2)
collect/display data that we collect while we have control of the machine
(our metric takes a few seconds and the community seems to want a metric
that runs for a few hours). (1) is obviously the priority and (2) is
currently done to some degree. Presenting data that can dovetail into a
centers analysis, where costs come in, is a useful service and we should
take care to collect and package the data well (eg, Jed's ideas below).
to weigh a machine against
> > it’s budget. A benchmark is meant to highlight an objective
> > measurement.
> I agree that cost or power efficiency should not be in the metric. But
> I don't think anyone in this thread was proposing that. Rather, we're
> talking about how to normalize scaling plots and it turns out there is
> no normalization that people are happy with. That's why I like
> circumventing the arbitrary choices by instead plotting performance
> versus time-to-solution. Not quite the same information, but the
> important trade-offs are easy to see and there are no hidden choices.
> HPGMG-Forum mailing list
> HPGMG-Forum at hpgmg.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the HPGMG-Forum