On November 19, 2003 03:51 am, Niklas Bergh wrote:
> > On November 18, 2003 08:52 am, Martin Stone Davis wrote:
> > > We already back-off due to QR, so we already have a way of telling
> > > when every node is going to be ready. �The issue here is
> >
> > adding that
> >
> > > back-off time to the estimate() so that we can tell if the best
> > > strategy for a given query is to wait for the node to not be backed
> > > off or to go with an available node.
> >
> > Thinking about it, I can see a fairly simple change to out
> > acceptance logic that will might help.  Let say we work like this.
> >
> > IF key is in the DS or Failure table        \\ sames as now
> >     respond with DNF or data
> > IF loadEstimate fails request                       \\ same as now
> >     respond with QR
> >
> > \\ probalistic QR disabled for now
> >
> > generate the routing object
> >
> > if the probability of finding a route using the top X nodes
> > (where              \\ we just look at pSF
> >             backed off nodes have a pSF of 1.0) is less
> > than 0.xx                   \\ no messages are sent
> >     respond with QR
> >
> > Reset the above object's getNextRoute to the top of the table
> > and proceed as we do now
> >
> > What this will do is allow us to accept requests that we
> > probably can route using a good
> > route.   We want to recheck the backoff time when we actually
> > route, maybe the backoff
> > expired while we tried another node.
> >
> > Given that we have lots more queries that we can process and
> > we need to reject some,
> > this should let us reject those we would route using 'bad'
> > nodes or would RNF.
> >
> > Comments?
>
> To me this is very similar to m0davis suggestion about the 'request
> bins'. The main difference is that your idea is to QR when the score is
> to low whilst m0davis idea is to prioritize the requests that we get
> high routing scores for over those which we get low scores for.

Its based very much on m0davis's idea.

> I prefer m0davis method since it matches up better with NGR. NGR learns
> times needed.. Not the ugly QR binary on/off stuff. As I stated in
> another mail.. NGR indeed seems to be able to (as it is intendedc to)
> utilize time consumed for different tasks to do a better job while it is
> completely unable to cope with QR:s.

Its basicly a subset of m0davis's idea.  With his alg we would accumulate
several request and route those with the highest routing scores.  It would
require some tricky coding to get right.  What I propose would be simple
to code.  You are going to have to do math depending on pSf/timeout for either
to work - as it stands now, just because a node is not backed off does not
mean it will accept a request.  pSF is the best idea we have about how
often it will accept.

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

Reply via email to