On Fri, Oct 31, 2003 at 08:23:06PM +0000, Richard Lamont wrote:
> On Friday 31 October 2003 19:03, Toad wrote:
> > On Fri, Oct 31, 2003 at 04:52:42PM +0000, Richard Lamont wrote:
> 
> > > Following the recent improvements, pSuccessRatio seems to have
> > > moved from around 2% to around 2.5% and stayed there.
> >
> > Unfortunately psuccess doesn't seem to actually mean anything. And it
> > won't, at least not until we get the stats sorted out - probably by
> > implementing the SFT...
> 
> Well, if the stats are telling porkies, diagnosing the problem is even
> more difficult.

Yeah. :(
> 
> > > The objective is to get keys from a node that has them to the node
> > > that wants them. Freenet needs designing to maximise the speed of
> > > delivery of keys.
> > >
> > > If messages take up none of the bandwidth, then no requests could
> > > be made, and no keys would get delivered.
> > >
> > > If messages take up all of the bandwidth, then there would be no
> > > bandwidth left for keys, and no keys would get delivered.
> > >
> > > Somewhere between these two extremes, there must be an optimum
> > > split of bandwidth between keys and messsages that maximises the
> > > rate at which keys are delivered.
> >
> > What would you suggest? Giving trailers a higher priority will only
> > drive up messageSendTimeRequest and make the node useless. Is it a
> > routing problem? Is it that 95% of requests are for data that simply
> > isn't on the network? In which case, we need to look at enlarging the
> > failure table - 1000 keys is simply too small for current usage
> > IMNSHO.
> 
> The usefulness of a node is a function of its ability to deliver keys.
> 
> All I'm suggesting is that if the bandwidth is allocated 80:20 keys:messages
> then Freenet will be four times as fast than if the ratio is the other
> way round (as it appears to be at present).

Well if you have any practicable ideas of how to achieve this, now's the
time for them.
> 
> As for which messages are dropped or delayed, I couldn't comment. I don't
> know enough about it, and AFAIK FNP isn't documented so it's more or less
> impossible to learn how it works. All I'm saying is that if 80% of the
> bandwidth is spent shifting keys, then Freenet will go four times as fast
> as at present.

There is a spec in CVS somewhere...
> 
> > > The ideal rate of sending messages would be the rate which is just
> > > enough to keep keys flowing fast enough to fill the pipe. Any more
> > > is worse than useless, because it will just waste bandwidth.
> >
> > Why is "wasting bandwidth" such a big deal?
> 
> I can't believe you are asking this question.
> 
> Most Freenet users are on cable modem or ADSL connections, with typically
> a 128 or 256 kbit/sec uplink. The default config bwlimits to 12000 Bytes/sec.
> If 80% of that is spent on message overhead, then payload can only move
> out of a node at a paltry 2400 bytes/sec. And even that is shared between
> all the TXing connections.
> 
> Bandwidth is the most precious and scarce resource that Freenet uses.
> Everything else - e.g. CPU and disk - is plentiful and cheap by comparison.
> So to optimise Freenet's performance, it is essential to use Bandwidth
> economically. Surely I can't be the first one to point out something
> as basic and obvious as this?
> 
> > > Have the developers given this any thought? All I've seen are one
> > > or two hints in various places that messages must have priority
> > > over payload. If developers are working on the basis of this
> > > belief, then that is the root of the performance problem.
> >
> > I repeat my question. There are lots of packets mainly because
> > routing doesn't usually find the node we want the first time, and
> > because a lot of queries fail, many of which are for data that wasn't
> > there in the first place. Messages must have priority over trailers
> > in order to be delivered in reasonable time - or that's the current
> > understanding anyway. QueryRejected messages and so on have a lower
> > priority though. But the basic misunderstanding here: we create
> > messages for a reason, we don't just pluck them out of thin air...
> 
> Please step back from all this detail and look at the big picture.
> 
> You create messages for a reason: to obtain keys. Keys can only flow
> at a certain rate, limited by bandwidth. There's no point in sending
> any more messages than just enough to get those keys sent, because they
> cannot achieve anything. 

Yes, but we have to FIND the keys in order to transfer them. They do not
come out of thin air.
> 
> As for which messages to throw away, I don't know enough about FNP's
> innards to offer any useful suggestion, and I'm more than happy
> to leave that to you.
> 
> I'm reminded of an old cartoon I saw years ago. It showed a huge open plan
> office with hundreds of men in suits sat at desks punching calculators.
> At the back the back of the room, there were two men watching this scene.
> One of them, the boss, says to the other, "Do you know, Jenkins, no matter
> how many accountants I employ, I can't seem to find out where all the
> money is going."
> 
> Freenet is a bit like that. It spends too much of the firm's money on
> accountants.
> 
> 
> -- 
> 
> Richard Lamont
> <[EMAIL PROTECTED]>
> OpenPGP Key ID: 5ABEC9C3  http://www.stonix.demon.co.uk/key.txt 
> Fingerprint: 9DEE 7113 DF02 A516 404C  22AC 1FF6 185D 5ABE C9C3

-- 
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