RFC: Incoming/outgoing discovery with fewer roundtrips / more bandwidth
Dirkjan Ochtman
dirkjan at ochtman.nl
Tue Sep 29 01:46:38 CDT 2009
On Tue, Sep 29, 2009 at 08:28, Peter Arrenbrecht
<peter.arrenbrecht at gmail.com> wrote:
> But I think I have an idea on how to solve all of the glaring problems
> I have so far discovered. It revolves around:
>
> * Taking samples along graph paths starting at the heads, not just
> from the somewhat arbitrary order of revs in a repo.
> * Doing this breadth-first, maybe with higher sample density near the
> repo heads, until we have gathered enough samples for this roundtrip.
> * Having, per sample, an upper bound on the lengths of the paths that
> go from this path towards tip or the next sample. This limits the
> amount of data we have to send for detailed pruning.
> * Using an iterative approach where the two parties exchange samples
> until all uncharted regions have been discovered sufficiently. Each
> sample may yield hints on how to restrict the range of further samples
> to only still uncharted regions. I expect that in most cases the
> initial sample will already do the trick.
Oh, and I think Benoit had some ideas for using an algorithm to
communicate the graph more efficiently, I think he used something
called dominators or dominance.
Cheers,
Dirkjan
More information about the Mercurial-devel
mailing list