<div dir="ltr"><div>Thanks everyone. I still think there is a even better phrase for this, like, static scaling? Because unlike strong/weak scaling, concurrency is fixed (hence "static") and we only scale the problem, so this is a mix between strong and weak scaling. </div><div><br></div><div>¯\_(ツ)_/¯<br><br></div>Anyway, what I really wanted to say is, it's good to know that these "dynamic range/performance spectrum/static scaling" plots are designed to go past the sweet spots. I also agree that it would be interesting to see a time vs dofs*iterations/time plot. Would it then also be useful to look at the step to setting up the preconditioner?<br><div><br></div><div>Justin</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Aug 30, 2016 at 1:55 PM, Jed Brown <span dir="ltr"><<a href="mailto:jed@jedbrown.org" target="_blank">jed@jedbrown.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">Mark Adams <<a href="mailto:mfadams@lbl.gov">mfadams@lbl.gov</a>> writes:<br>
> I would guess it is the latter.<br>
<br>
</span>In this case, definitely.<br>
<span class=""><br>
> It is hard to get "rollover" to the right. You could get it on KNL<br>
> (cache configuration of HBM) when you spill out of HBM.<br>
<br>
</span>Yes, but the same occurs if you start repeatedly spilling from some<br>
level of cache, which can happen even if the overall data structure is<br>
much larger than cache. Not all algorithms have the flexibility to<br>
choose tile sizes independently from problem size and specification;<br>
it's easy to forget that this luxury is not universal when focusing on<br>
dense linear algebra, for example.<br>
<span class=""><br>
> Personally, if you are you going to go into this much detail (eg, more than<br>
> just one plot) I would show a plot of iteration count vs problem size, and<br>
> be done with it, and then fix the iteration count for the weak scaling and<br>
> dynamic range plot (I agree we could use a better name).<br>
<br>
</span>Alternatively, plot the performance spectrum (dynamic range) for the<br>
end-to-end solve and per iteration. The end user ultimately doesn't<br>
care about the cost per iteration (and it's meaningless when comparing<br>
to an algorithm that converges differently), so I'd prefer that the<br>
spectrum for the end-to-end application always be shown.<br>
</blockquote></div><br></div>