[HPGMG Forum] Kiviat diagrams
John Shalf
jshalf at lbl.gov
Tue May 27 23:04:43 UTC 2014
Yes,
I think we are in agreement here Jed. The FOM is orthogonal to the Kiviat diagrams.
I think that is the point I’m trying to make. The Kiviat just makes sure you have a balanced benchmark.
That is different than having said benchmark optimize for an FOM.
This behavior should be expected since the FOM is extensive and the Kiviat is ideally scale-free (which is the opposite of extensive).
So my final point is that you should not worry about the fact that the Kiviat diagram will not necessarily diagnose a benchmark that is performing poorly on your FOM (that is not its purpose… its not even designed for that).
-john
On May 27, 2014, at 9:15 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> Mark Adams <mfadams at lbl.gov> writes:
>>> "challenging" parts of the hardware architecture. Having the most
>>> efficient DOF/s for this particular solver is orthogonal. You want to
>>> focus on mathematical rigor (in the same way that Linpack has a
>>> mathematically rigorous definition) and exercising the hardware aspects
>>> equally.
>>>
>>>
>> One of my narratives (see my slides) is that we are staying faithful to the
>> HPL model by: 1) solving a hard problem (fully connected global problem
>> with known mathematical complexity), 2) using the best method to solve it,
>> 3) using the best computer science to implement it.
>
> I just want to note that once we define the FOM, the code (and
> configuration) will be optimized for the FOM. Kiviat diagrams for
> code/configuration that does not maximize the FOM only describe the path
> you might see from algorithms doing something similar in a less
> optimized way. (This might look familiar because production apps
> typically have less focused optimization effort than benchmarks.)
More information about the HPGMG-Forum
mailing list