From rusty at rustcorp.com.au Sun Sep 27 04:24:45 2015 From: rusty at rustcorp.com.au (Rusty Russell) Date: Sun, 27 Sep 2015 13:54:45 +0930 Subject: [Lightning-dev] Ionization Protocol: Flood Routing In-Reply-To: <20150925050836.GA11549@navy> References: <8761314tpn.fsf@rustcorp.com.au> <38D4A6A1-474D-49A4-8380-76B65AFBEB70@gmail.com> <87a8sbs5t1.fsf@rustcorp.com.au> <20150925050836.GA11549@navy> Message-ID: <87bncoqyk2.fsf@rustcorp.com.au> Anthony Towns writes: > On Fri, Sep 25, 2015 at 09:56:02AM +0930, Rusty Russell wrote: >> [ Pieter Wuille Cc'd for pubkey recovery, search for "recovered" ] >> Mats Jerratsch writes: >> >> Indeed. Random selection helps, here, but analysis will be interesting. >> > How have you ended up with a number of beacons you need? 12 seems so >> > low, I can?t imagine so few nodes to support all transacting for even >> > 10 minutes.. >> As we keep the last 100 sets of beacons, the load is spread a little. > > Does that actually work? Old beacons don't do any good if the payee > doesn't use them when advertising a route; but old beacons also don't > get their fee updates propogated, and aren't known by people who only > just joined the network. I don't think you could usefully keep more than > the last 2 or 3 sets of beacons? I was thinking they all count as active. No point turning over all the beacons every block. >> To start, I was thinking you establish channels with 5 random nodes. > > I think Barabasi-Albert graphs are probably pretty reasonable here -- > you start by establishing channels to N nodes, selected randomly but > favouring nodes in proportion to how connected they already are. > > https://en.wikipedia.org/wiki/Barab%C3%A1si%E2%80%93Albert_model Interesting. If we assume growth only (as I suggested) we end up with a geometric distribution (as backed by Section 5 of http://www.e-publications.org/ims/submission/index.php/BEJ/user/submissionFile/10315?confirm=c40442a0 ). It's not clear to me (without reading the paper) how much attachment bias is required for a power law to apply, however. Larger nodes may be incentivized (through profit) to have better uptime, leading to a slight attachment bias. Definitely second order. It's not clear that we want a power law: do we? Cheers, Rusty.