Notes

Outline

Users, Pricing and Resource Reservation: Managing Expectations. Jon Crowcroft, http://www.cs.ucl.ac.uk/staff/jon/ jon@cs.ucl.ac.uk

Two Rival Approaches Reservations Call,signal,qos route, route pin, admission control, call log Hop by hop Accounting Principal objective metrics: User:Call blocking prob. Net: Erlangs Adaption Congestion signal (loss, ecn, delay) Feedback control loop TCP eqn etc etc Principle objective metrics: User throughput Utilisation

M3I/JISC Approach End system based adaption – components: Network distributes congestion data via ECN marks Network Monitoring Agents distribute general load information to Tariff Agents Tariff Agents distribute current spot price per region by multicasting region id (AS#?) plus price per kbps per rtt End user or risk broker choose rate End user may pay broker a long term price or choose to adapt to spot prince Metering and Adaption models are distributed signed and sealed to end users, who send franked packets using them (otherwise packets are just BE)

Partial Deployment… This can be partially deployed (piecewise) as it only needs congestion point to distribute infromation, and to be monitored. Doesn’t require unloaded cores to do anything Doesn’t usually require (most) edge routers to do anything End system module can be plugin

Hypothesis…part 1 Internet Traffic Matrix Is inherently unpredictable How bad do users feel about unpredictable call blocking?resource reservation must FAIL (i.e. you must have a non zero call blocking probability) or it is pointless. If it fails, you have to look at user disatisfaction with call blocking If the user satisfaction with call blocking is bad enough, you will lose customers to someone else who provisions. So the REAL argument is what are the relative costs of call blocking versus provisioning.

Hypothesis …part deux How bad do users feel about unpredictable pricing? In between these extremes there is dynamic pricing - you can offer users a spot price instead of a futures price - users dont like this much (we have done some experiments...) but sometimes they'll buy into it - Then adapt the user’s behaviour instead of the application’s behaviour - whether you can do this enough to make the mean to peak ratio low enough without incurring the wrath of users; (unpredictable price is nearly as bad as unpredictable call success/blocking probablity) is the slightly more variable version of problem 1.

Experimental Plan… UK Academic Community has 1 major bottleneck: UK-US link (only OC-12J Can meter; can place proxy servers at each end; can distribute price based on metering to proxies, and distribute applets which adpt to end users (at US and UK end) Have captive groups lined up…

What Next? Need to look at budgets Need to look at deployment Need to look at protection domains too (confidence/.ring fenced qos etc) Would like to do differentiation (esp. of delay bounded/EF like services( within same experiment!.

Any Questions… What would you pay? When? How (standing order, direct debit, e-cash etc etc) To whom ? (can middleware make money)? J L?