[Yt-dev] new parallelHF load-balancing
Stephen Skory
stephenskory at yahoo.com
Mon May 10 15:07:56 PDT 2010
Hi all,
I've just pushed to devel a new load-balancing implementation for parallelHF. The old way would repeatedly read data off disk during load-balancing. Instead, this one uses a random subset of all the particles and does the load-balancing in memory on the subset, only. This should be substantially faster, especially on slow disks. It turns out that using a very small percentage of the total number of particles can achieve decent load-balancing.
There is a new parameter that controls how many particles are used for load-balacing, called 'sample'. It gives the fraction of the total number of particles to be used for balancing. The default is 0.03, which means 3% of the total are used. I've experimented down to 0.3% and seen reasonable results there, too.
h = yparallelHF(pf, threshold=160.0, safety=1.5, \
dm_only=False,resize=True, fancy_padding=True, rearrange=True, sample=0.03)
Let me know if you run into any problems using this. I've done some testing and the results are identical to the old method, but something could be wrong, of course.
_______________________________________________________
sskory at physics.ucsd.edu o__ Stephen Skory
http://physics.ucsd.edu/~sskory/ _.>/ _Graduate Student
________________________________(_)_\(_)_______________
More information about the yt-dev
mailing list