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.
¯\_(ツ)_/¯ 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? Justin On Tue, Aug 30, 2016 at 1:55 PM, Jed Brown <[email protected]> wrote: > Mark Adams <[email protected]> writes: > > I would guess it is the latter. > > In this case, definitely. > > > It is hard to get "rollover" to the right. You could get it on KNL > > (cache configuration of HBM) when you spill out of HBM. > > Yes, but the same occurs if you start repeatedly spilling from some > level of cache, which can happen even if the overall data structure is > much larger than cache. Not all algorithms have the flexibility to > choose tile sizes independently from problem size and specification; > it's easy to forget that this luxury is not universal when focusing on > dense linear algebra, for example. > > > Personally, if you are you going to go into this much detail (eg, more > than > > just one plot) I would show a plot of iteration count vs problem size, > and > > be done with it, and then fix the iteration count for the weak scaling > and > > dynamic range plot (I agree we could use a better name). > > Alternatively, plot the performance spectrum (dynamic range) for the > end-to-end solve and per iteration. The end user ultimately doesn't > care about the cost per iteration (and it's meaningless when comparing > to an algorithm that converges differently), so I'd prefer that the > spectrum for the end-to-end application always be shown. >
