Re: Impact of window size on delta performance

> Hi. A couple of times I've said we are performing 35 times worse than
> Phong Vo's vcdiff implementation and I didn't know why. Well, I now
> think that we're only doing that badly on some fairly brutal inputs,
> and that we can do better on such inputs by increasing our window size
> a little bit.

I had similar results, yes.

> Phong Vo claimed to get a 273K diff in about 44 seconds; he didn't say
> what hardware he was using or what kind of windowing his program uses.
> You can see that we actually do substantially better on diff size than
> his program did if we use a window size of 64K or greater.

Don't worry about the times right now. We can always speed things up
later on, if necessary (and -- Measure First�!).

> Still, I'm going to increase the default window size to 100K. A
> comparison of two similar tarfiles represents a kind of worst
> reasonable case, and while I'm sure I could get a better picture of
> that case by comparing other combinations of tarfiles, I don't want to
> do that much of a research project.

That's reasonable. 16k was a ridiculously small amount, anyway; just
something to get started with.

-- 
Brane �ibej
  home:   <brane_at_xbc.nu>            http://www.xbc.nu/brane/
   ACM:   <brane_at_acm.org>           http://www.acm.org/