Why should we use delay-based congestion control? Delay-based Congestion Control for Low Latency The delay signal is: I I I David Hayes <[email protected]> David Ros <[email protected]> I requires no special signal May also aid in using available capacity more quickly Ideally: 25 - 26 September 2013, London I This work has been funded by the European Community under its Seventh Framework Programme through the Reducing Internet Transport Latency (RITE) project (ICT-317700) ¯ RITE) David Hayes & David Ros (UiO & IMT w A measure of the quantity we are trying to reduce A fundamental characteristic of packet switched networks DBCC for Low Latency 25–26 September 2013, London 1/6 A measurement based CC that tries to keep latency low ∴ congestion control for low latency, SHOULD include delay as A congestion measure. ¯ RITE) David Hayes & David Ros (UiO & IMT w DBCC for Low Latency 25–26 September 2013, London 2/6 Why is delay a difficult signal to use? Delay Measures 0.115 0.11 Measures: 0.105 I TRTT I TOWD I usually interested in Tq , but requires TRTT_min or TOWD_min I but with multi-path TOWD or TRTT maybe just as important as Tq I gradient or change in delay I ... 0.1 0.095 0.09 340 350 360 370 380 390 400 410 420 430 time (s) OWD signal Dealing with Noise: ¯ RITE) David Hayes & David Ros (UiO & IMT w DBCC for Low Latency 25–26 September 2013, London 3/6 Raw / probabilistic – collective behaviour matters I Filtering: mean, median, exponential, kalman ¯ RITE) David Hayes & David Ros (UiO & IMT w DBCC for Low Latency 25–26 September 2013, London 4/6 Delay based CC for: Loss based (LB) congestion control and delay based (DB) congestion control do not coexist well I LB tries to fill the queues I DB tries to keeps them small I homogeneous systems (e.g. data centres) I Scavenger class traffic I small buffers incentives for deploying DBCC I Hybrid two-state methods I only reduces delay in the long term I Reverse path (RTT) I A way forward Coexistence? I I I I I (as loss based flows disappear) long term is still important! perhaps policing isolation of delay-based flows in combination with other signals (such as ECN) DBCC is not THE solution, but should be PART of the solution ¯ RITE) David Hayes & David Ros (UiO & IMT w DBCC for Low Latency 25–26 September 2013, London 5/6 ¯ RITE) David Hayes & David Ros (UiO & IMT w DBCC for Low Latency 25–26 September 2013, London 6/6