1. © British Telecommunications plc 1 nice traffic management without new protocols Bob Briscoe Chief Researcher, BT Oct 2012

2. © British Telecommunications plc 2 vertical stripes: this season’s colour increasingly access no longer the bottleneck –PON, FTTP, FTTdp, FTTC –bottleneck moving deeper –becoming similar to campus LAN, data centre each customer’s average access utilisation v low –1-3% average during peak hour (some 100%, but rarely at the same time) –if provision bottleneck for the worst case load seen, it leaves a lot of leeway for much worse cases traditionally the bottleneck solves this with –(weighted) fair-queuing / round robin* about isolation from random bad performance –not about skimping on capacity time rate time rate * or policed per-customer b/w limits (for higher utilisation customers e.g. DC) viewed at second mile

3. © British Telecommunications plc 3 enforces 1/N shares* so that’s fine? No –when average N is so low –a few more long-running customers than planned can increase N significantly –thereby greatly decreasing everyone else’s 1/N share the problem: –N depends heavily on presence of high util’n customers –usually few, but when many, service seems crap –large buffers make this much worse but not the only problem 1/N is ‘fair’ at each instant, but not over time fair queuing: so 1990s time rate * as does WRR (and TCP sort-of) time rate

4. © British Telecommunications plc 4 ConEx: so y=2014 y = y(now) + 2 the ConEx rationale FQ + volume limits? –hi vol customer only a problem if with other hi vol cust’s Lack of complete solution led to non-neutral solutions, then... Comcast fair-share –limit highest volume customer(s) only if cable congested –better, but penalises hi-vol even if transport yields [LEDBAT] research goal: we ain’t seen nothing yet on the Internet...if we designed the network for un-lame transports ConEx rationale is actually in two parts: 1.rationale for using congestion-volume as a metriccongestion-volume 2.need a change to IP (ConEx) to see congestion-volume Is there an 80% solution without changing IP?

5. © British Telecommunications plc 5 simpler to code than draw bottleneck congestion policer: next season’s colour foreach pkt { i = classify_user(pkt) d i += w i *(t now -t i ) //fill t i = t now d i -= s * p //drain if (d i <0) {drop(pkt)} } s : packet size p : drop prob of AQM backhaul &accessnetwork policer incoming packet stream meter w1w1 w2w2 wiwi … d i (t) cici congestion token bucket Y FIFO buffer p(t) AQM … outgoing packet stream

6. © British Telecommunications plc 6 bottleneck congestion policer (BCP): features predictable quality for the many –keeps queue v short –by focusing discard on those who most push at the queue* tends to WRR if customer traffic becomes continuous app-neutral applicability –same as any per-customer limiter –state: per-customer configured allowance and usage-level –drop-in to: BRAS / MSE, RNC, OLT, DC access –few simultaneous customers (or many) where bottleneck location varies, still need to evolve to ConEx typical anomalous * relative to their allowance to do so

7. © British Telecommunications plc 7 next steps (next season’s shoes) plan to open-source BCP –not yet fully agreed internally baby step –gets industry used to congestion-volume as the metric time rate time rate

8. © British Telecommunications plc 8 nice traffic management without new protocols Q&A discussion spare slide

9. © British Telecommunications plc 9 measuring contribution to congestion = bytes weighted by congestion level = bytes dropped (or ECN-marked) = ‘congestion-volume’  marginal cost of capacity as simple to measure as volume bit-rate time congestion time 10GB 0.01% congestion 1MB 1% congestion 1MB 300MB 100MB 3MB 1% 0.01%

10. © British Telecommunications plc 10 wiwi d i (t) cici kw i C AQM p(t) policer main congestion token bucket meter congestion burst limiter police if either bucket empties actually each bucket needs to be two buckets to limit bursts of congestion