Hi Matt,

Looks like some good speedup:
Old: 5.53041e+01
New: 2.57021e+01

This was on a 6 level AMR everywhere.  I will note that as you mentioned, the disk cache does do a lot, as here was the results from the first run:
Old: 5.51695e+01
New: 2.23177e+02


On Fri, Aug 27, 2010 at 1:06 PM, Matthew Turk <matthewturk@gmail.com> wrote:
Hi everyone,

I added the quad tree projection to the tip of the hg repository.

If you could please put it through a few paces, that'd be great.  You
can access it as the attribute "quad_proj" on a hierarchy.  I've done
some speed testing and the generality necessary to get it to work with
sources and whatnot makes it a bit slower than if it was just fed the
data, but it's still about twice as fast on my datasets as the old
projection method.

It doesn't work in parallel yet, but give it a shot on some big
datasets anyway -- I tested it on a 1024^3 L7 and it worked fine in
serial.  But, I know how to parallelize it, and the parallelization
process should also speed up the serial projections.  (It will allow
us to better choose the batching of IO.)  I think this projection
method should also scale *much* better for parallelism, because we
shouldn't have to do domain decomposition to parallelize, which was
always a killer on nested grid simulations.

I'd like to push forward with this, so if you get a chance to try it,
report back with any issues.  I'm sure there will be several...

I've placed a comparison test here:


(But note that disk cache will favor the old projection method the
first time you run, so run at least twice.  :)

Yt-dev mailing list

Samuel W. Skillman
DOE Computational Science Graduate Fellow
Center for Astrophysics and Space Astronomy
University of Colorado at Boulder