On Thursday 31 January 2008 16:06, Michael Rogers wrote:
> On Jan 30 2008, Matthew Toseland wrote:
> >So I propose to implement weighted-coin-followed-by-HTL. This should cause 
> >very little disruption, as we won't have very short requests, in fact the 
> >code changes would be very minor.
> 
> Sounds good - your arguments against parallel/repeated inserts are 
> convincing. What about automatic re-requests - if they're triggered by a 
> timeout or a DNF, could an attacker measure the fraction of re-requests 
> that reach his node and work out the likely distance to the originator? 
> Would randomising the re-request interval help?

At the moment, they are handled at the client layer. In fproxy for example we 
always try 3 times - and we start the second request immediately after the 
first one fails, if there is nothing else running. It might be possible to do 
some stats, although I dunno exactly how; the second request would be routed 
differently with per-node failure tables.
> 
> >We can always change it later on to for 
> >example pure weighted coin (although I'm not convinced that will work well 
> >for inserts).
> 
> Me neither - although we could use a lower pDrop for inserts if many of 
> them are failing to reach their targets.

Well for inserts subject to FEC, weighted coin might work. The concern is that 
not all inserts are subject to FEC - for example the top block of a 
splitfile. Frost posts too, but then they'll be quite popular and well 
propagated through ULPRs.
> 
> Alternatively, we could set pDrop very low (say 1%) but toss the coin 
> *before* trying each peer, which would count loops and overloads. If the 
> topology's any good, most searches will get close to the target and then 
> bounce around getting RejectedLoops until the coin comes up tails, which is 
> fairly cheap. But if the topology's broken and there are a lot of RNFs, 
> we'll have a higher chance of escaping dead ends.

Also means that if there's a lot of load we visit fewer nodes per request. I 
dunno, sounds a bit messy, we'd need to simulate it. OTOH it'd be great to 
have something which leaks as little info as possible.
> 
> Cheers,
> Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL: 
<https://emu.freenetproject.org/pipermail/devl/attachments/20080131/0bf5a865/attachment.pgp>

Reply via email to