Notes

Outline

Explicit Charging for Multicast (ECMmmm...) Jon Crowcroft, UCL@AdastralPark.biz http://www.cs.ucl.ac.uk/staff/jon/ jon@cs.ucl.ac.uk

IP Multicast evolution in IETF SSM now flavour of month (lots of ISPs “about to deploy”) sing source matches most compelling applications extant (multisource apps can use multiple ssm groups, or a server (hopefully fault tolerant/dual redundent - e.g. game) Works well with interdomain (lose BGMP/MASC rubbish:-) and with emerging RMT solutions

Multicast Transport Evolution in IETF RTP/RTSP done deal RMT: PGMcc + ALC sender adaption exactly as TCP receiver adaption by selection of layer(s) Both use TCP eqn model for rate => can replace loss signal with ECN

Two Rival Approaches Sender Pays Why multicast (scaling video servers)? Why help ISP? What about receiver (ac)counting? What about unknown receiver set sizes/distributions/distances? Receiver Pays What about benefit to other users? What about benefit to ISP? What about late joiners? etc etc

Tristan Henderson’s Approach Suggest we model it as a mix Sender pays based on ingress rate and ... Receivers pay based in share - Estimate #of receivers periodically Period is based on statistics of type of session (have lots of studies of games, movies, music “holding times”) http://www.cs.ucl.ac.uk/staff/t.henderson Assume there’s a pay per content too, so sender already has feedback channel to count receivers  - could be RTCP based or PGMcc or PIM-SSM ...

A Mark, a Buck, a Frank or a Yen? Correct model for ECN setting for multicast packet? IP Level (ECN bis) ECN is set (typically) by RED (c..f AF work by Dina Papagiannaki at UCL/Sprint..) Equivalent for Multicast is to set ECN at router at ingress to congested link, then: randomly select which branch to forward the packet with ECN (i.e. clear for all other branches)

Aha, Aha, Aha....Aha Correct model for reacting to ECN in a multicast packet? RMT Level (ECN ter) Single rate (PGMcc) Sender adaption - requires feedback to sender - EXACTLY as TCP - each receiver gets marks and wraps them around - will find same operating point as PGMcc with loss feedback Multirate (ALC) receiver adaption - each receiver chooses layers based on ECN

...In a Rich Man’s World so above scheme gives proportional fairness... can use M3I model for pay per ECN unmodified notice that the overall traffic adaption rate is slower But its slower for n unicasts anyhow, so? includes a two-level model (content and network costs) but no coupling (necessary-  unless you want to bundle) - just rx counting, which ISPs wanted anyhow, and CDNs have to have

Any Questions… What would you pay? notice cost may not be predictable - not just coz of congestion price, but now share (risk is other people game:-) No problem: have a risk broker CDNP may carry risk (has statistics on “popularity” of game,film,song...tristan’s stats... J L?