[HPGMG Forum] More thoughts on the Kaviats
jed at jedbrown.org
Mon Jun 9 13:30:55 UTC 2014
Theodore Omtzigt <theo at stillwater-sc.com> writes:
> The derivative indeed is the quest, but don't get too comfortable with
> that either. Because we are talking about a discrete event system, there
> is no linearity to speak off. If you deplete a particular resource, your
> performance will fall dramatically ruining your derivative.
For a task that does only one thing, we might expect to have only one
active constraint. But these apps and benchmarks (apart from STREAM) do
several different things, so overall performance is a sum of the active
constraints of the parts. It's true that the derivative is not
continuous, but for any given position, we should see that more than one
machine attribute modulating performance. Collecting these data for
several of today's machines would be valuable.
> I think the solution lies in the 'big data' that will be generated by
> running the benchmarks on hundreds or thousands of configurations. Weak
> scaling measurements are the best in discovering the cliffs in the
> resource availability, so instead of asking the hardware or a simulator
> to cripple the resource, ask the application to find the resource
> bottleneck and hammer it.
Unfortunately, this is complicated and hard to do in a systematic way
across a range of applications. For our purposes, I think we need
something simpler to measure.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 818 bytes
Desc: not available
More information about the HPGMG-Forum