TCP Friendliness Considered un-friendly
| Jon Crowcroft | |
| WAGN |
| The Internet “Community” has adopted TCP-like behaviour as a Gold Standard: | |
| This talk is about why this may be very detrimental to future choice | |
| We’ll use examples from transport protocol requirements | |
| Other arguments are also possible | |
| What is TCP like behaviour exactly? | |
| What other requirements might there be on transport protocols? | |
| What impact (if any) does this have on IP (routers)? | |
| Are there other implications? |
| Basic rule is that on an end-to-end time scale, a flow should try to achieve fairness (e.g. Raj Jain’s fairness definition) with TCP – two approaches are | ||
| implement AIMD | ||
| Use TCP equation (1/rtt*sqrt(loss) | ||
| Has been made “respectable” by Kelly, Maulloo & Tan | |
| Proportional fairness includes link (if 1/RTT == link cost) as well as max-min share of pipe over timescale of adaptation… | |
| Has a bug: needs user choice of route over short timescale… | |
| Has a missing middle (risk?) | |
| Multicast | |
| Mobile | |
| Link cost varies (CDMA) | |
| Striped flows | |
| etc | |
| Microsoft use a 1.2Gbps link 24*365 | |
| Jim Gemmell computes if they multicast each thing once, and used PGMcc (or RLC) they’d only need a 10kbps link | |
| BUT your download would take 1 yearJ |
| Sweet spot is a function of receivers timescales/needs and the number of receivers | |
| Non trivial rate selection | |
| Multi-rate is better but not ideal – still complex | |
| Need to consider sender, and network topology has least implicitly |
| If I move, the link share with other people varies | |
| Why should my rate vary? | |
| Why should theirs? | |
| What if my path is predictable? |
| CDMA nets share is affected directly but other users use looks like noise | |
| What if I do multi-hop radio net, but can dynamically choose size of cell/hop? | |
| What if I can choose whether to be a router or not dynamically? |
| Sometimes my access link is fast (e.g. GRID) 1.2Gpbs, but each path on a route is slower | |
| What if I Stripe data | |
| Much the same as the layered multicast case, only perhaps a tad easier… |
Impact on IP forwarding/route?
| Local FIFO+drop or ECN gives a shadow packet (or bit if virtual buffer) price | |
| TCP equation. is proportional fairness | |
| Can in principle translate all the previous problems into a form of this | |
| But users may not want to (c.f. audio/video adapt price, not rate? What about session level price stability preferences?) |
| Need to consider enforcement – M3I model is to do this via money and admission (per packet, per flow, per aggregate) – maybe other permits could be allowed | |
| Need to implement hierarchical monitoring | |
| Risks need understanding |
| In practice, provisioning enforces capacity rules in most ISPs (tier 1 core’s over provisioned compared to access ISPs access links, but access ISPs are 3-1 to7-1) | |
| In practice, scheduling & queue management in the net are pretty chaotic…at the whim of vendor/provider |
| Need to reshape the agenda | |
| Hard to do (the IETF is dominated by dogma just like ISO and ITU were 10-20 years ago) | |
| But hey, we are researchers, so just do it! |
| M3I has a metering/edge admission architecture | |
| Need to quantify set of costs and distribute them from within net | |
| Need to talk to router people about deploying smarter ECN like stuff | |
| Need a break… |