Re: [lmi] Benchmarking: gcc-8 beats gcc-10 soundly?

[Greg Chicares]

On 2020-09-20 12:21, Vadim Zeitlin wrote:
> On Sun, 20 Sep 2020 09:59:31 +0000 Greg Chicares <gchicares@sbcglobal.net> 
> wrote:
[...]
>  It's still rather strange that Wine influences the run time of mostly
> numeric code so much, but maybe there are some calls to MSW APIs still done
> even in this case.

'wine' version 4 compared to version 5:

  https://lists.nongnu.org/archive/html/lmi/2020-09/msg00024.html
| transplanted third scenario is much worse (166175 vs 102521),
| transplanted second scenario is much better (26650 vs 33180).

The one that's better with wine-4.0.3 ("second scenario") is
the one with no solves, so its total time might be thought of as
  1 * startup + 1 * calculations
whereas the second might be
  1 * startup + 50 * calculations
leading to a system of equations like
  166175 = s0 + 50 * c0
   26650 = s0 +  1 * c0
  102521 = s1 + 50 * c1
   33180 = s1 +  1 * c1
with this solution:
  s0 = 23802.55
  s1 = 31764.88
  c0 = 2847.45
  c1 = 1415.12
i.e.
  wine-4  wine-5
   23803   31766  fixed startup
    2847    1415  monthly calculations, per iteration
from which I can conclude...nothing.

I was kind of hoping that "per iteration" would be the
same while "fixed startup" would differ, in which case
perhaps wine's disk reads might have gotten much slower.
Or the opposite, which might have suggested a different
conjecture that we could test.

>  Anyhow, I think it's still worth comparing the binaries produced with 8.3
> and 10.2 under native MSW. If there is no difference there, we probably
> should just stop there and chalk all this up to Wine. If there is still a
> difference even there, we need to continue with investigating it as
> discussed before.

We're in full agreement.

[I could send you a couple of fardels, but trading binaries
is a security concern, and it'll only take a couple minutes
for you to generate them yourself.]