[petsc-users] PetscLogFlop for a sqrt()

Matthew Knepley knepley at gmail.com
Tue Apr 21 12:29:15 CDT 2015

On Tue, Apr 21, 2015 at 12:23 PM, Justin Chang <jychang48 at gmail.com> wrote:

> Last question
> I would like to report the efficiency of my code. That is, flops/s over
> the theoretical peak performance (on n-cores). Where the TPP is clock *
> FLOPS/cycle * n. My current machine is a Intel® Core™ i7-4790 CPU @ 3.60GHz
> and I am assuming that the FLOPS/cycle is 4.
> One of my serial test runs has achieved a FLOPS/s of 2.01e+09, which
> translates to an efficiency of almost 14%. I know these are crude
> measurements but would these manual flop counts be appropriate for this
> kind of measurement? Or would hardware counts from PAPI?

1) For this, I think the manual counts are good enough for the estimate

2) You really should not compare to TPP, which ignores memory bandwidth


You should run the STREAMS benchmark on your machine as the link says, and
use a "roofline" model to estimate the peak performance based on bandwidth.
Here is
a talk on that model:


and here is a paper which does exactly this for sparse MatVec (Krylov


Basically, you have some smaller multiplier to the bandwidth (arithmetic
intensity), which
gives you the real performance upper bound, not TPP.



> Thanks,
> Justin
> On Tue, Apr 21, 2015 at 11:16 AM, Jed Brown <jed at jedbrown.org> wrote:
>> Matthew Knepley <knepley at gmail.com> writes:
>> > Flop is Floating Point Operation. The index calculation is an Integer
>> > Operation. I agree that we could probably start counting
>> > those as well since in some sorts of applications its important, but
>> right
>> > now we don't.
>> Index calculations often satisfy recurrences that the compiler folds
>> into pointer increments and the like.  Also, some architectures, like
>> PowerPC, have floating point instructions that include mutating index
>> operations in the true spirit of RISC. ;-)

What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150421/5c37119c/attachment.html>

More information about the petsc-users mailing list