[yt-dev] [yt_analysis/yt] Clump finding doesn't use chunks (issue #546)

[Matthew Turk]

Hi Chris,

I figured I'd reply to the mailing list since I think this is a big
issue.  I'm not sure that I have a good plan in place for clump
finding in 3.0, which I think is probably overall quite broken.  The
issue of the IsBound quantity is that it has never ever worked in
parallel, and I am not surprised it does not now.  We have in the past
mandated that some of the quantities -- notably this one -- be
calculated in serial and with a single array of the entire dataset.  I
believe returning to that, where we chunk "all" instead of "io", is
necessary.  My suggestion is to examine the force_unlazy property and
change the chunk if need be.

I've been doing a lot of thinking about how to port clumps to 3.0, and
I think what we should investigate is reworking clump finding entirely
to use a forest of disjoint-sets.  I implemented a first pass at this
about two years ago, but wasn't able to finish or work out the kinks.
Another option would be to create join/split contour trees, which
would enable single-pass construction of dendograms.  There are some
other ways we could explore this, but I think it might be a very fun
problem that requires a whole lot of time...

-Matt

On Mon, Apr 8, 2013 at 4:48 PM, Christopher Moody
<issues-reply at bitbucket.org> wrote:
> New issue 546: Clump finding doesn't use chunks
> https://bitbucket.org/yt_analysis/yt/issue/546/clump-finding-doesnt-use-chunks
>
> Christopher Moody:
>
> Clump-finding is broken for the chunk system; currently IsBound returns pot/kin for every chunk. This fails spectacularly when the chunking goes level-by-level, where we only include cells at the same levels in the potential calculation. Instead, it should add every chunk to a persistent octree, and in the combine step it should calculate the binding energy.
>
>
> Responsible: juxtaposicion
> --
>
> This is an issue notification from bitbucket.org. You are receiving
> this either because you are the owner of the issue, or you are
> following the issue.
> _______________________________________________
> yt-dev mailing list
> yt-dev at lists.spacepope.org
> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org