[GRASS-dev] Re: parallelizing GRASS modules

[Hamish]

Michael:
> So improvements to the segment library should
> help both
[r.cost and r.walk]

Perhaps better stated as learning how to run the
segment library in parallel. The changes would be
in the modules AFAIU, not in the library per se.
(although perhaps the could be, or both the lib and
module working together for it, ??)


> (and also v.surf.bspline and v.surf.rst??)

v.surf.bspline does use GRASS's segment library,
but v.surf.rst quadtree segmentation is something
different.

note Soeren has recently parallelized the bit of
the gmath library which v.surf.bspline uses which
for me made it 3.5x faster on 6-cores, but incurred
a 40% overhead penalty. Hopefully by parallelizing
the segments instead of the inner loops of the
gmath library we can get that overhead down near 0%
of the overall task.


also to consider is that the segment library was
created mainly as a way to keep RAM usage down on
big jobs. running all segments at once negates that
advantage. (but of course you can limit the number
of threads launched at run time using environment
variables if that is a problem)


Hamish