I'll try to re-answer some of the ones that weren't
clearly answered before. Keep in mind I'm new here
too.
> * How are queries routed *alternatively* in order
> to attempt to bypass a Data
> Not Found ("DNF") event? (I'm especially
> interested in this question, as
> I haven't figured out how I want to do it in Mnet
> yet. ;-))
According the original freenet doc (there was a nice
extened one in HTML that I cann' find anymore), if a
node gets a DNF it forwards to it's second chioce. In
doing so it takes the HTL coming back in the DNF
message to keep things from recursing indefinately.
Think of the freenet as a graph and you're doing a
depth first traversal with HTL number of nodes being
accessed. In short it sends the message to its next
best choice.
I was considering proposing an alternative to this,
which might allow for some limited parallel branching.
That's a post in its own right.
> One potential problem with the way it does that is
> that it rests on the
> assumption that true DNFs -- queries for data that
> is absent from Freenet --
> is evenly spread throughout the keyspace. I can
> imagine cases where that
> assumption becomes untrue, even for extended periods
> of time or for
> significant fractions of the keyspace.
Yes, I can think of two.
1) Your specialized in a very small part of the
keyspace and key that has been lost, suddenly becomes
popular again and therfore is requested repeatly just
inside your area.
2) Some aversary wants to damage some keyspace so he
sends bogus requests for data there. This could
discredit servers in that specialization.
> Obviously this can work only because the job of
> partitioning the key space is
> separated from the which-peer-to-route-to question
> in Mnet. (That job will be
> performed, in Mnet v0.7, by a separate mechanism as
> part of the DHT-like
> routing.)
DHT eh... I've sometimes wondered if we chould just
fix the routing using one of the more rigorous
algorithims like CAN, Chord, or DHT. The reason we
don't is to try to build security into the routing.
I'm not sure the current design really does this
anyway though.
I'm almost wondering if one can prove that single
completely untrusted nodes can't build such a network.
Maybe I should check your project some.
This isn't necessarily bad, but it was a little
surprising to me when
> I realized it, and it is a way in which NGrouting is
not scale-free.
> That is: suppose there is a small Freenet where one
node receives
> requests which are evenly distributed across 0.4 of
the keyspace.
> Suppose there is a large Freenet where one node
receives requests
> which are evenly distributed across 0.004 of the
keyspace. The former
> will move 5 of the points after a few dozen new
samples have come in.
> The latter will move only 2.
I thought this too at first glance but no. What will
happen is the first two points will come in toward the
middle of the hump. Once they get into the .004 of
the keyspace, queries just outside those two points
will yank the others in one by one.
Only problem and this counts for all of these learning
sytems: the uncertainty priciple. When we measure how
well another node is in a certain keyspace, by giving
it queries, we affect what we're messuring.
__________________________________________________________________
Gesendet von Yahoo! Mail - http://mail.yahoo.de
Logos und Klingelt�ne f�rs Handy bei http://sms.yahoo.de
_______________________________________________
devl mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
