[petsc-users] strong-scaling vs weak-scaling

[Mark Adams]

>
>
>> I have seen some people claim that strong-scaling is harder to achieve
>> than weak scaling (e.g., https://www.sharcnet.ca
>> /help/index.php/Measuring_Parallel_Scaling_Performance) and generally
>> speaking it makes sense - communication overhead increases with concurrency.
>>
>
I would back up and avoid this sort of competitive thing. Weak and strong
scaling are different, but valid metrics. They each tell you different
things.  Are oranges better than apples? Depends but they are both useful
and of the same basic class.

Just to add, I've started to like Jed's metric of "dynamic range" (for lack
of a better word). This is like strong scaling except instead of fixing the
problem and increasing the parallelism you fix the parallelism and decrease
the problem size. Then plot this with Time vs. N/Time (rate). This has the
nice property, like weak scaling, a horizontal line is perfect. The
rollover point is the lower bound on turnaround time (I call these
turnaround time plots sometimes), and both axis are interesting (size, P or
N, is not interesting, but rate and time are). But, this is just
different.  Because it keeps parallelism the same, your latencies are a
constant (ignoring algorithmic effects of problem size), so in a sense it
is between strong and weak scaling.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20160822/25585314/attachment.html>