[Haskell-cafe] Problem with benchmarking FFI calls with Criterion

[Janek S.]

I tested the same code on my second machine - Debian Squeeze (kernel 2.6.32) with GHC 7.4.1 - and 
the results are extremely surprising. At first I was unable to reproduce the problem and got 
consistent runtimes of about 107us:

benchmarking FFI/C binding
mean: 107.3837 us, lb 107.2013 us, ub 107.5862 us, ci 0.950
std dev: 983.6046 ns, lb 822.6750 ns, ub 1.292724 us, ci 0.950

benchmarking FFI/C binding
mean: 108.1152 us, lb 107.9457 us, ub 108.3052 us, ci 0.950
std dev: 916.2469 ns, lb 793.1004 ns, ub 1.122127 us, ci 0.950

I started experimenting with the vector size and after bumping its size to 32K elements I started 
getting this:

benchmarking FFI/C binding
mean: 38.50100 us, lb 36.71525 us, ub 46.87665 us, ci 0.950
std dev: 16.93131 us, lb 1.033678 us, ub 40.23900 us, ci 0.950
found 6 outliers among 100 samples (6.0%)
  3 (3.0%) low mild
  3 (3.0%) high severe
variance introduced by outliers: 98.921%
variance is severely inflated by outliers

benchmarking FFI/C binding
mean: 209.9733 us, lb 209.5316 us, ub 210.4680 us, ci 0.950
std dev: 2.401398 us, lb 2.052981 us, ub 2.889688 us, ci 0.950

First result is always about 39us (2,5 faster, despite longer signal!) while the remaining 
benchmarks take almost two times longer.

Janek


Dnia niedziela, 25 listopada 2012, Janek S. napisał:
> Well, it seems that this only happens on my machine. I will try to test
> this code on different computer and see if I can reproduce it.
>
> I don't think using existing vector is a good idea - it would make the code
> impure.
>
> Janek
>
> Dnia sobota, 24 listopada 2012, Branimir Maksimovic napisał:
> > I don't see such behavior neither.ubuntu 12.10, ghc 7.4.2.
> > Perhaps this has to do with how malloc allocates /cachebehavior. If you
> > try not to allocate array rather use existing one perhaps there would be
> > no inconsistency?It looks to me that's about CPU cache performance.
> > Branimir
> >
> > > I'm using GHC 7.4.2 on x86_64 openSUSE Linux, kernel 2.6.37.6.
> > >
> > > Janek
> > >
> > > Dnia piątek, 23 listopada 2012, Edward Z. Yang napisał:
> > > > Running the sample code on GHC 7.4.2, I don't see the "one
> > > > fast, rest slow" behavior.  What version of GHC are you running?
> > > >
> > > > Edward
> > > >
> > > > Excerpts from Janek S.'s message of Fri Nov 23 13:42:03 -0500 2012:
> > > > > > What happens if you do the benchmark without unsafePerformIO
> > > > > > involved?
> > > > >
> > > > > I removed unsafePerformIO, changed copy to have type Vector Double
> > > > > -> IO (Vector Double) and modified benchmarks like this:
> > > > >
> > > > > bench "C binding" $ whnfIO (copy signal)
> > > > >
> > > > > I see no difference - one benchmark runs fast, remaining ones run
> > > > > slow.
> > > > >
> > > > > Janek
> > > > >
> > > > > > Excerpts from Janek S.'s message of Fri Nov 23 10:44:15 -0500 2012:
> > > > > > > I am using Criterion library to benchmark C code called via FFI
> > > > > > > bindings and I've ran into a problem that looks like a bug.
> > > > > > >
> > > > > > > The first benchmark that uses FFI runs correctly, but
> > > > > > > subsequent benchmarks run much longer. I created demo code
> > > > > > > (about 50 lines, available at github:
> > > > > > > https://gist.github.com/4135698 ) in which C function copies a
> > > > > > > vector of doubles. I benchmark that function a couple of times.
> > > > > > > First run results in avarage time of about 17us, subsequent
> > > > > > > runs take about 45us. In my real code additional time was about
> > > > > > > 15us and it seemed to be a constant factor, not relative to
> > > > > > > "correct" run time. The surprising thing is that if my C
> > > > > > > function only allocates memory and does no copying:
> > > > > > >
> > > > > > > double* c_copy( double* inArr, int arrLen ) {
> > > > > > >   double* outArr = malloc( arrLen * sizeof( double ) );
> > > > > > >
> > > > > > >   return outArr;
> > > > > > > }
> > > > > > >
> > > > > > > then all is well - all runs take similar amount of time. I also
> > > > > > > noticed that sometimes in my demo code all runs take about
> > > > > > > 45us, but this does not seem to happen in my real code - first
> > > > > > > run is always shorter.
> > > > > > >
> > > > > > > Does anyone have an idea what is going on?
> > > > > > >
> > > > > > > Janek
> > >
> > > _______________________________________________
> > > Haskell-Cafe mailing list
> > > Haskell-Cafe at haskell.org
> > > http://www.haskell.org/mailman/listinfo/haskell-cafe
>
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://www.haskell.org/mailman/listinfo/haskell-cafe