On Mon, Oct 13, 2003 at 12:34:21PM +0200, Some Guy wrote:
> --- Todd Walton <[EMAIL PROTECTED]> wrote: 
> > On Sun, 12 Oct 2003, Benjamin Coates wrote:
> > 
> > > A discussion on #freenet leads me to suggesting this change to QueryReject 
> > > behavior:
> > >
> > > The goal here is to reduce the message traffic to overloaded nodes without 
> > > seriously changing network behavior.
> > 
> > What advantage does this have over the way it already happens, i.e. 
> > NGRouting.  With NGRouting, if a node is overloaded and QRs, everybody 
> > adjusts their appraisal of the node downwards, and message traffic to that 
> > overloaded node is reduced.
> > 
> > What advantage does your above idea offer over the NGRouting effect?
> > 
> > (Not that there are no advantages, just asking what they are.)
> Be Nice.
> 
> NGR and the old routing take a long time (hours) to train up.  So, they would also 
> take a while to
> "train down".  This is probably too slow to react to some overload problems.

NGR should react pretty quickly to QueryRejects. It's not an estimator
it's just a decaying average.

> 
> You may also argue that the rejections caused by overload as noise, which if treated 
> the same as
> normal DNFs muck up the picture and make it that much harder for NGR or LGR (last 
> generation
> routing) to learn.
> 
> For these reasons, it seems like you'd want to have special DNFs to let overloaded 
> nodes say hey
> "leave me alone for a bit".
> 
> That said I do understand why it seems cleaner to let NGR handle the problem.
> ---------
> My suggestion: Look at how Ethernet worked on coax.  You'd broadcast your message 
> and if you got a
> collision, back off for a random period and try again.  For every repeated failure 
> you'd double
> the wait time until some high limit.
> 
> It can be shown that such a simple algorithm can allow a theoretically infinite 
> number of devices
> to get an equal share of the media, without the collisions wasting more than a 
> particular fraction
> of the bandwidth (I think <70% for ethernet).
> 
> The problem is similar with an overloaded freenet node:
> We have a limited resource.  
> We don't know who all is using it.
> If we all jump on it, it won't be able to service any of us, because all upbandwidth 
> will be used
> in rejecting messages.
> 
> Maybe if all the nodes would back-off exponentially on overload DNFs, we wouldn't 
> get those 16.88%
> accepted queries Mathew reffered to in his next post.

-- 
Matthew J Toseland - [EMAIL PROTECTED]
Freenet Project Official Codemonkey - http://freenetproject.org/
ICTHUS - Nothing is impossible. Our Boss says so.

Attachment: signature.asc
Description: Digital signature

_______________________________________________
Devl mailing list
[EMAIL PROTECTED]
http://dodo.freenetproject.org/cgi-bin/mailman/listinfo/devl

Reply via email to