Quoth Stefan Sperling <stsp@stsp.name>:
> On Thu, Feb 10, 2022 at 12:57:14AM -0500, ori@eigenstate.org wrote:
> > Playing with it, if I reduce the minimum chunk
> > size to 32, the pack gets significantly smaller:
> > 
> > 	1.6188G	eecede5b54cafd33515f01101c139b098370ec3a.pack
> > 
> > And switching from sha1 to murmurhash2 for the delta
> > hash table is a significant speedup.  It drops the time
> > to repack the plan9front repo from 67 seconds to 48
> > seconds.
> 
> Thank you, Ori!  Is the patch below equivalent to what you did?
> 
> I am sorry I did not find time yet to provide you with numbers that
> show our current performance with respect to size and time.
> I could still do that if it would still be useful to you.
> 

Yes, that's about right.

The smaller the chunk size, the more 'seeds'
we get to expand matching deltas from, which
is more computation but also better matching.

Tweaking the split mask is also probably
worth playing around with, but I haven't
done much with it. Smaller masks means
more candidates for chunk boundaries;
it controls the average chunk size.