[yt-dev] FLASH data in 2.x versus 3.0

[Matthew Turk]

Hi all,

Because of some testing/modification for the pull request that Andrew
submitted, I tested some IO speeds for FLASH, which I decided to share
here.  What it came down to was testing a full dataset read, with
slicing, versus constructing the H5S and only reading the 8x8 slab in
question.  I ran this with 30 iterations on the cluster merger FLASH
dataset.

Constructing a slice and only reading that in 2.x: 1.212s per slice
Reading the full block dataset and discarding all but the slice of
interest: 1.283s per slice
In 3.0: 0.4856s per slice

Mind you, seems like we have a big performance regression for finding
the location of the most dense point, but that should be easily
fixable.  The vast majority of the savings come from pushing the
selection of points *inside* the IO function, which lets us retain the
H5D handles in between IO calls.

-Matt