[Haskell-cafe] Re[2]: Haskell performance in heavy numerical computations

[Bulat Ziganshin]

Hello Simon,

Monday, July 10, 2006, 1:12:13 PM, you wrote:

>> numerical speed is poor in ghc 6.4, according to my tests. it's 10-20
>> times worse than of gcc. afair, the mandelbrot benchmark of Great
>> Language Shootout proves this - despite all optimization attempts,
>> GHC entry is still 5-10 times slower than gcc/ocaml/clean ones

> We average 1.3x slower than C in the shootout. Thanks to Don Stewart for
> the following stats...

as i once wrote, most of shootout benchmarks depends on the libs. for
example, multi-threading benchmark is much faster on GHC than on GCC
because former has user-level threads support pre-included and for
later the external libs should be used (testing policy prohibits using
of additional libs)

> The numerical floating-point-intensive benchmarks:

>      mandelbrot  2.7x C    (Clean 1.7x, OCaml 2.4x, C++ 2.6x)
>      n-body      2.1x C    (Clean 0.9x, OCaml 1.3x, C++ 1.4x)

that is the benchmarks that i had in mind

> http://shootout.alioth.debian.org/gp4/benchmark.php?test=mandelbrot&lang=all

the same benchmark with N=600 and Sempron processor:

http://shootout.alioth.debian.org/debian/benchmark.php?test=mandelbrot&lang=all

here GHC is 10x slower than GCC and more than 5 times compared to
Clean and Ocaml. i think that this is the real computing speed
difference while with N=2000 C/Ocaml/Clean programs was really limited
by memory speed.

at least the same situation was for trivial "a[i]=b[i]+c[i]" benchmark
discussed half-year ago - when arrays was small enough to fit in cpu
cache, speed difference was 20 times. on large arrays C program was
limited by memory speed and speed difference was only 3 times


>   *  2.5x C
>   *  3.2x C
>   * 11.0x C
>   * 22.0x C

>   * Known, or expected, to improve under Data.ByteString

one more sign that this is more testing of libraries than compilers. for
example, revcomp test performs only about 1000 operations in main
language but calls regex functions with 100kb strings. not strange
that in this test TCL is now fastest entry

about numerical apps - afair, Manuel Chakravarty works on new version
of his parallel arrays engine


-- 
Best regards,
 Bulat                            mailto:Bulat.Ziganshin at gmail.com