Not that I'm aware of. The current plan is to see if it appears under a
network running Opennet (path folding) with added churn (nodes joining for
a while taking a randomized position, participating some in swapping, then
disappearing for good).

Simulating a darknet-style network about if this is happening is more
complicated, I don't know of a straightforward way to simulate churn where
nodes leave for good in a darknet and with new ones arriving being placed
well in the topology (with no path folding).

Vilhelm

On Wed, Aug 15, 2007 at 12:14:16PM -0500, Ian Clarke wrote:
> Indeed, have we actually replicated the clustering effect with churn in a
> simulation?
> Ian.
> 
> On 8/14/07, Matthew Toseland <toad at amphibian.dyndns.org> wrote:
> >
> > The reasons we need periodic randomization are 1) churn results in
> > clustering
> > and 2) attacks. So it would be interesting to look at it with some churn.
> >
> > I'll have a look at the rest tomorrow.
> >
> > On Tuesday 14 August 2007 21:50, vive wrote:
> > > Here are some experiments on running the swapping algorithm with
> > periodic
> > > randomization of node positions. The idea is to study measures against
> > > position clustering, and later on what can be made to counter attacks
> > > against the swap algorithm. No attacks have been simulated yet, this is
> > the
> > > pure impact of randomizing positions on the swapping algorithm.
> > >
> > > The network size was fixed to 10000 nodes, HTL to 100, and the average
> > > degree was varied. Each run a network was initiated as the Kleinberg
> > model,
> > > measured by routing a while on it, and then positions were uniformly
> > > randomly assigned between 0 and 1 before staring the swapping algorithm
> > to
> > > recover the efficient routing.
> > >
> > > The major thing to study was the frequency of randomizing the position.
> > One
> > > run was evaluated for each different frequency X. Randomizing frequency
> > X
> > > was used as follows: each node randomizes its location each X'th swap
> > that
> > > goes out from that node. To avoid cyclic behaviour (at least of the
> > simple
> > > case, perhaps this needs to be extended) each nodes counter for doing
> > this
> > > is initialized independenty between 0 and X before starting the
> > > simulations. Lets say a node randomly initalizes its counter to y
> > (between
> > > 0 and X), then it will randomize its location on the swap
> > > y,y+X,y+2X,y+3X... that starts from it with a random walk (of length 6).
> > > This is an easy way to let a node stay with a location for a while
> > (unless
> > > swapping) and to let the neighbors route through it before randomizing
> > > again. X=0 corresponds to not randomizing at all.
> > >
> > > Discussing the results:
> > > There seems to be room for running a low-frequent randomization to begin
> > > with. For one case (average degree=10) it even seems to perform better
> > in
> > > some cases with randomization with a quite large period (this surely
> > > depends on that randomization can help the algorithm get out of some
> > > suboptimal configurations). This does not seem to happen for a larger
> > > average degree, perhaps because each randomization will have impact on
> > more
> > > links (but that is no controlled answer, just speculation at the
> > moment).
> > >
> > > Another note: how quickly the algorithm improves the network from the
> > > entirely randomized state is not of foremost interest here, the
> > interesting
> > > thing is what happens when the network is rather stable from swapping
> > for a
> > > long period. Therefore its the resulting level (on the right sides of
> > the
> > > plots) that are important, but the levels seem to have settled around
> > some
> > > performance (I will rerun the first simulation again to see how pure
> > > swapping without no randmization behaves in a longer run)
> > >
> > > Finally, we dont know if the observed behaviour here will be the same in
> > > the current Freenet topology (since the efficiency of the swapping
> > > algorithm depends on how much the topology deviates from the Kleinberg
> > > model in the first case). Therefore any implementation attempt should be
> > > using defensively low periods of randomization.
> > >
> > > regards,
> > > /Vilhelm
> >
> > _______________________________________________
> > Tech mailing list
> > Tech at freenetproject.org
> > http://emu.freenetproject.org/cgi-bin/mailman/listinfo/tech
> >
> >
> 
> 
> -- 
> Founder and CEO, Thoof Inc
> Email: ian at thoof.com
> Office: +1 512 524 8934 x 100
> Cell: +1 512 422 3588
> AIM: ian.clarke at mac.com
> Skype: sanity

> _______________________________________________
> Tech mailing list
> Tech at freenetproject.org
> http://emu.freenetproject.org/cgi-bin/mailman/listinfo/tech
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
URL: 
<https://emu.freenetproject.org/pipermail/tech/attachments/20070815/0de65120/attachment.pgp>

Reply via email to