[erlang-questions] Strange optimization result

[Anders Nygren]

On 10/21/07, Steve Vinoski <vinoski@REDACTED> wrote:
> On 21 Oct 2007 11:41:12 +0200, Bjorn Gustavsson <bjorn@REDACTED>
> wrote:
> > "Anders Nygren" <anders.nygren@REDACTED> writes:
> > >
> > > My only idea is that this is caused by some pathological CPU cache
> > > behaviour. Does anyone have a better explanation?
> >
> > Could be. I have another GUESS: using a tuple instead of dict will
> > result in a smaller heap size for the spawned process, which will lead
> > to more frequent garbage collections. You could use the spawn_opt()
> > BIF with the min_heap_size option.
> >
> > To get a similar heap size as when using a dict, use
> >
> >         erts_debug:size(wfbm3:init()).
> >
> > which is 1120.
> >
>
> I didn't try this or Bob's suggestion of simulating a tuple, but I did
> eliminate a dict lookup in the code that finds shift values, and also
> changed the macros to use hard-coded string lengths (since the strings are
> fixed) rather than recalculating them with length/1, and got another 25%
> speedup:
>
> <http://steve.vinoski.net/blog/2007/10/21/faster-wf-still/>
>
>

So I then took Steve's latest and
- changed to using a tuple instead of dict and VICTORY :)


Steve's tbray16 on my laptop
real    0m10.929s
user    0m18.477s
sys     0m0.560s

With tuple instead of dict
real    0m14.126s
user    0m22.373s
sys     0m1.924s


- added [{min_heap_size, 3000}] to the spawn, as suggested by Bjorn

With tuple and min_heap_size, 1000
real    0m7.358s
user    0m11.725s
sys     0m0.480s

With tuple and min_heap_size, 2000
real    0m6.760s
user    0m10.957s
sys     0m0.388s

With tuple and min_heap_size, 3000
real    0m6.305s
user    0m10.149s
sys     0m0.380s

After that increasing the heap size does not make any difference.

So my next step will be to play with garbage collection and see what I can get.

/Anders