Include Benny's improved interpolation patch

[wd15]

Not sure where the patch is. I'm assuming Jon has it.

Imported from trac ticket #348, created by wd15 on 05-25-2011 at 11:38, last modified: 09-01-2011 at 14:57

[wd15]

guyer attached nearest2.png on 09-01-2011 at 14:26

[guyer]

Replying to [wd15](issue #252):

Not sure where the patch is. I'm assuming Jon has it.

Posted to the mailing list at http://article.gmane.org/gmane.comp.python.fipy/2033

Trac comment by guyer on 08-30-2011 at 14:13

[guyer]

Fixed in r4742

Merged Benny's code, generalizing the determination of closest points in numerix.nearest().

It turns out that the chunk size does not benefit from being very big, and can be catastrophic if it's too big (horrendously slow disk thrashing if swap space is available or crashing with Memory Error if it's not).

10⁸ bytes seems to get most of the benefit without the risk on a variety of architectures. See calculation in trunk/fipy/tools/numerix.py@4742#L470. Note that disk thrashing seems to set in well below the available RAM.

• (# points) x (# cells) gives (# comparisons)
• time x (clock speed) gives cycles
• (# cells) x (chunk size) x dimensions x (bytes / value) gives (bytes / chunk)

Trac comment by guyer on 09-01-2011 at 14:57