> On Oct 10, 2016, at 7:09 PM, Jed Brown <j...@jedbrown.org> wrote:
>
> Barry Smith <bsm...@mcs.anl.gov> writes:
>
>> Funny, I thought Jed said I stripped out all the PAPI stuff? Maybe I
>> should really strip it out, has not been used in years and as Jed
>> says it has Jack's slimeprints on it.
>
> I haven't touched it.
I know you didn't touch it. You said I (meaning Barry) stripped it out.
> Looks like you added it in 2009. I think we had
> a mailing list thread on it around that time.
>
> Counting flops with hardware counters is really error-prone on current
> hardware. I don't think we should ever trust it for linear algebra. It
> would perhaps be handy for people implementing their physics and
> discretizations, but the numbers may vary widely based on compiler/flag
> choices.
Independent of its reliability, I found the way of managing what was
reported through its API in 2009 was very cumbersome and thus just left what
little I had done in place and never pursued it.
As I said in my other email I'm fine with PETSc's current logging that gives
one the "big picture" of what is taking time and what is doing well and poorly
and think one can just use external tools like vtune for focused studies on
kernels people are interested in. I got some push back from others in follow up
email; but didn't see any rational for the push back besides "it would be nice
if we had more logging..." :-( I think people may underestimate the amount of
work and maintenance needed for PETSc to properly manage more logging then we
already do (there is already more lines of convoluted logging code then I would
ideally like to have).
Barry
