[HPGMG Forum] Do we want the benchmark to go into intrinsics?

Mark Adams mfadams at lbl.gov
Tue Apr 29 17:56:27 UTC 2014


On Tue, Apr 29, 2014 at 1:28 PM, Sam Williams <swwilliams at lbl.gov> wrote:

> There's also a deeper question on how you evaluate/rank.
>
> I was thinking of requiring submitters to submit 6 numbers at a given
> concurrency.
> - reference implementation for N, N/8, and N/64 DOF/process
>

I still think that trying to define optimized, and therefore not in the
reference implementation, is just asking for trouble.

I would have to think that Erich would not like this.  Have you run it by
him?  Perhaps I am being paranoid and conservative.


> - full fury implementation for N, N/8, and N/64 DOF/process
> Ranking would be based on a linear combination of these (the first three
> could be weighted by 0)
>
>
A concern that I have your proposal and Jed's is that the community might
not like it. Or it is a bit distraction until we get a seat at the table.
 I envision getting support from the community and having them run and
analyze whatever you want to put in the repo.  Jed has a "sample" parameter
that runs jobs at multiple scales that he is probably proposing as evolving
into a benchmark that includes dynamic range, like yours.  I suggest that
we be vague about this.  Just say that it is good to look at dynamic range
and this is raw data on it.  And have the code produce DOF/process at N for
optimized version, as well as other data if desired.  If people ask "what
is the metric", then I like saying "DOF/sec, just like HPL".

Additionally, I have been saying the metric is DOF/sec to draw a contrast
with HPCG.  HPCG does not make it clear that they are not solving the
equations and I want to highlight this distinction.  I will tone that down
and add that we are "exploring dynamic range metrics".

Sam, it sounds like you have ideas for dynamic range and so we will not try
to make that common in HPGMG at this point.  I was thinking that we would
try to make this common as I had not heard that you have ideas for this.
 Jed has thoughts on this issue and you can both explore them.  You and Jed
can each explore these ideas and get feedback from users and managers like
Erich and Horst.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://hpgmg.org/lists/archives/hpgmg-forum/attachments/20140429/a01e0c7b/attachment.html>


More information about the HPGMG-Forum mailing list