On Tue, Jun 25, 2002 at 01:18:37AM -0400, Gianni Johansson wrote:
> On Monday 24 June 2002 23:19, Matthew wrote:
> 
> > > On Tue, Jun 25, 2002 at 04:16:49AM +0100, Matthew Toseland wrote:
> > > Patch committed, changes overload handling to:
I hoped it might get your attention :)
> > > always accept if in failure table (as now)
> > > 60%-70% load:
> > > accept HTL=1, HTL=0 (this is after we have decremented it)
> > > accept if in most successful 1/256th of the keyspace (by % requests
> > > fulfilled successfully)
> See below
> 
> 
> > > accept if in datastore
> > > 70-80% load:
> > > accept HTL=0
> > > accept if in datastore
> > > 80%+ load:
> > > reject all except in fail table
> I'm glad to see someone actively working on fixing the overload problem.
> 
> Unfortunately, I don't think this approach to selecting which requests to 
> acccept is a good idea.
> 
> Here are my concerns:
> 
> 1) Your approach will penalize the nodes that route well (i.e. route to a 
> node that has a high chance of having the data, without consuming much htl) 
> at the expense of those that route poorly (i.e. nodes that route to many 
> nodes unsuccessfully before landing at your node with a low htl).  I don't 
> see how this can be good for the global (network wide) routing effectiveness 
> / specialization.
> 
> It would be a disaster if everyone setup their nodes to route every request 
> to every other node they could get a noderef for with an htl of 0 or 1.
Okay... this is a critique of the thing Ian forwarded, which I didn't
see at the time. You want to never prioritise low HTLs?
> 
> 2) Also, I don't think that its a good idea to try to pick a winning discrete 
> segment of the keyspace and ignore all others.  Real nodes don't have 
> successful request key distributions with single peaks. 
> 
> What I proposed was that each request would  be accepted or qr'd based on its 
> approximate Ps(k) value. 
> 
> What I imagined is that the implementation would keep a Psmin minimum 
> acceptable Ps(k) value which it increases under overload and decreases when 
> not overloaded.  All requests with Ps(k)  > Psmin would be accepted.  
OK. How do we determine Psmin? This was a hack, you understand, I'd like
to implement the Psmin thing.
> 
> 
> 
> 3) I think you need to distinguish between successful  DataRequests and 
> InsertRequests.  When the network is performance is bad there will be a 
> disproportiate number of successful InsertRequests as compared to the number 
> of successful DataRequests.  InsertRequests just mean that data was inserted 
> -- not that it was inserted where anyone else will ever find it.
> 
> I think that it would better to do the inbound request triaging based on 
> DataRequests only.
Hmmmm.
> 
> 
> Regards,
> 
> --gj
> 
> 
> p.s. Don't take that proposal as gospel.  I haven't heard any comment on it 
> from Oskar yet.  It has not yet sustained the withering critique typical of 
> the devl mailing list ;-).

Attachment: msg03334/pgp00000.pgp
Description: PGP signature

Reply via email to