From: Jan Svitok <jan.svitok@...>
Date: 2007-03-12T07:37:51+09:00
Subject: Re: Trying to make Array#collect massively parallel with OpenMP

On 3/11/07, Daniel Berger <djberg96@gmail.com> wrote:
> Hi all,
>
> Windows XP Home
> VC++ 8 (free edition)
>
> Just for kicks I tried creating a parallel Array#collect method, which
> I called Array#acollect (asynch. collect). I added the following C
> code to array.c, rebuilt and reinstalled, but it doesn't seem to be
> any faster. Could this be an issue with my compiler? Or a Windows
> thing?
>
> static VALUE rb_ary_acollect(VALUE ary){
>    long i;
>    VALUE collect;
>
>    if (!rb_block_given_p())
>            return rb_ary_new4(RARRAY(ary)->len, RARRAY(ary)->ptr);
>
>    collect = rb_ary_new2(RARRAY(ary)->len);
>
>    #pragma omp parallel for
>       for (i = 0; i < RARRAY(ary)->len; i++)
>          rb_ary_push(collect, rb_yield(RARRAY(ary)->ptr[i]));
>
>    return collect;
> }
>
> rb_define_method(rb_cArray, "acollect", rb_ary_acollect, 0);
>
> # bench_collect.rb
> require "benchmark"
>
> MAX = 4000
>
> array = []
> MAX.times{ |n|
>    array[n] = 2 * n
> }
>
> # No significant difference (?)
> Benchmark.bm(30) do |x|
>    x.report("Array#collect"){
>       MAX.times{ array.collect{ |e| e += 4 } }
>    }
>    x.report("Array#acollect"){
>       MAX.times{ array.acollect{ |e| e += 4 } }
>    }
> end
>
> Ideas?
>
> Thanks,
>
> Dan

Hi,

I'm no expert in either ruby internals or openmp. I've just a few
ideas for you (though most of them will be obvious probably):

1. http://msdn2.microsoft.com/en-us/library/fw509c3b(VS.80).aspx says
you need to add /openmp compiler switch (try using _OPENMP define to
see if the compiler recognizes omp pragmas). #include "omp.h" might
help as well.

2. in the same page they say you won't see any difference if the whole
loop runs under 15 ms on a specific machine (i.e. the thread startup
time)

3. try running one loop outside the benchmark to setup the thread pool

4. is the assignment intentional (e += 4)?

5. (just for my curiosity:) are you using 1.8 or 1.9?

6. I've heard that 1.8 interpreter runs in one thred. I suppose your
code runs correctly with multiple threads because there should not be
any (re)allocations (e.g. it might crash if there was a local var in
the block). Right?

7. what machine are you running this code on? (if you send me the
binary or patch I might try it on my core2 machine if that helps)

8. I suppose the difference might be bigger if you used more
complicated (longer) block (relates to #2)

9. if everything fails, try running pure c loops (without calling ruby
functions) with omp optimisation, possibly wrapped into a ruby
function to see if omp makes at least difference at c level

ok, I've run out of ideas for now...

Jano