Incentive-Compatible Differentiated Scheduling Background 2 - PowerPoint PPT Presentation

Incentive-Compatible Differentiated Scheduling Background 2 HotNets IV - November 2005 Scheduler Model 4 Implementation 6 Evaluation 8 Discussion 12 Wrap Up 13 Martin Karsten, Yunfeng Lin, Kate Larson School of Computer Science, University of Waterloo 200 University Ave W Waterloo, ON N2L 3G1 Canada hotnets2005_talk.fm Nov 14, 2005 1/14

Background Topic: Network Quality of Service Rate Control... • simple (edge) with rate-neutral FIFO scheduling → FIFO Principle ...vs. Delay Control Background 2 • priority scheduling → preferred service class Scheduler Model 4 • allocation-based scheduling Implementation 6 ⇒ Multi-class Admission Control → Complicated! Evaluation 8 Discussion 12 Wrap Up 13 ICDS: Reconciliation of Delay Control and FIFO Principle • rate control oblivious to delay control hotnets2005_talk.fm Nov 14, 2005 2/14

Background (cont’d) Alternative Motivation: Queueing Delay • ...produced by buffering • ...required for bursty traffic → Fate-sharing between bursty and smooth traffic? Background 2 Typical "Internet Applications" Scheduler Model 4 • varying flexibility of handling different rates Implementation 6 • some network loss tolerance Evaluation 8 • limited number of delay targets Discussion 12 • e.g. interactive human users Wrap Up 13 • for different media types hotnets2005_talk.fm Nov 14, 2005 3/14

Scheduler Model 0. Basics • ICDS provides n service classes with fixed delay targets 1. FIFO Principle • relative service rate = relative arrival rate • at time t : arrival rates a , link capacity C → compute service rate r Background 2 ( ) a i t ( ) Scheduler Model 4 r i t C ∑ = - - - - - - - - - - - - - - - - - - - ( ) a j t Implementation 6 Evaluation 8 Discussion 12 2. Delay “Guarantee” → Packet Discard Wrap Up 13 • discard packets that cannot be forwarded in due time • non-trivial for varying rate allocation... hotnets2005_talk.fm Nov 14, 2005 4/14

Scheduler Model - Game-theoretic Properties Game • each player (traffic source) has fixed delay target • each player selfishly chooses service class Assumptions 1. lower delay ⇒ higher drop rate 2. delay exceeds target ⇒ zero utility Background 2 3. any delay lower than target ⇒ same utility Scheduler Model 4 4. lower drop rate ⇒ higher utility Implementation 6 5. service rate (throughput) unaffected by choice of service class Evaluation 8 Discussion 12 Wrap Up 13 Result: ICDS is strategy-proof • best strategy is to always choose true delay target (that is: highest delay lower than target) hotnets2005_talk.fm Nov 14, 2005 5/14

Implementation Overview packet discard Background 2 Scheduler Model 4 Implementation 6 classification rate-proportional Evaluation 8 scheduler packet queue Discussion 12 Wrap Up 13 rate estimation hotnets2005_talk.fm Nov 14, 2005 6/14

Implementation Details Rate Estimation • avoid arbitrary division → modify Time Sliding Window (TSW) • direct relative estimation: operate on arrived bytes rather than time Packet Scheduling • limited number of classes: scheduler no big concern? • prototype uses WF 2 Q+ Background 2 Scheduler Model 4 Implementation 6 Packet Discard Evaluation 8 • drop on departure? may not be efficient Discussion 12 Wrap Up 13 Rate Allocation and Delay • loose delay mode: ignore estimation errors and rate variation • introduces errors • strict delay mode: account for rate variation • check sum of rates against budget • implement rate increase immediately • implement rate reduction only after previous packets are served • conservative scheme → reduced resource (buffer) utilization hotnets2005_talk.fm Nov 14, 2005 7/14

Evaluation Simulation Experiment • dumbbell topology with 155 Mbit/sec at bottleneck • end-to-end latency: 30 msec → 60 msec round-trip latency • 3 traffic sources • CBR - 1 flow UDP/CBR with 15.5 Mbit/sec (10%) • TCP - 100 flows TCP/Greedy • Bursty - 32 flows UDP/Pareto with 93 Mbit/sec average rate (60%) • FIFO: 60 msec buffer Background 2 • ICDS: 3 delay classes Scheduler Model 4 • 10 msec Implementation 6 • 30 msec Evaluation 8 • 60 msec Discussion 12 • ICDS loose-delay mode ⇒ occasional delay violations Wrap Up 13 hotnets2005_talk.fm Nov 14, 2005 8/14

Evaluation (cont’d) FIFO with 60 msec buffer CBR TCP 140 Bursty 120 Background 2 Throughput (Mbit/s) 100 Scheduler Model 4 Implementation 6 80 Evaluation 8 Discussion 12 60 Wrap Up 13 40 20 0 0 10 20 30 40 50 60 70 80 90 100 Time (s) hotnets2005_talk.fm Nov 14, 2005 9/14

Evaluation (cont’d) ICDS with CBR in 10, TCP in 30, and Bursty in 60 msec class CBR TCP 140 Bursty 120 Background 2 Throughput (Mbit/s) 100 Scheduler Model 4 Implementation 6 80 Evaluation 8 Discussion 12 60 Wrap Up 13 40 20 0 0 10 20 30 40 50 60 70 80 90 100 Time (s) hotnets2005_talk.fm Nov 14, 2005 10/14

Evaluation (cont’d) Average Throughput in Mbit/sec Scenario (CBR/TCP/Bursty) CBR TCP Bursty FIFO (60/60/60) 13.6 44.5 75.9 ICDS (10/30/60) 13.5 45.2 72.4 ICDS (10/10/60) 14.1 34.5 75.1 ICDS (10/30/30) 13.7 33.0 72.0 Background 2 ICDS (10/60/60) 13.6 42.6 75.3 Scheduler Model 4 Implementation 6 ICDS (10/30/10) 12.4 50.5 59.2 Evaluation 8 Discussion 12 Wrap Up 13 • ICDS (10/30/60) provides “best” performance • “cheating” does not help • TCP can be affected by competing traffic - see ICDS (10/30/30) • no gain for Bursty → denial-of-service only • TCP target not obvious - compare ICDS (10/30/30) with ICDS (10/60/60) hotnets2005_talk.fm Nov 14, 2005 11/14

Discussion Essence of ICDS • proper incentives for burst control and/or traffic shaping • policy-free delay differentiation • no more fate-sharing for smooth and bursty traffic Deployment Scenarios • isolated deployment: delay differentiation without control regime Background 2 • overloaded nodes without sophisticated traffic management Scheduler Model 4 • e.g. peering exchanges? Implementation 6 • end-to-end rate control Evaluation 8 • domain deployment: admission control at edge gateways Discussion 12 • no static resource partitioning Wrap Up 13 • no signalling with internal nodes • multiple bottlenecks: no pay-bursts-once principle Traffic Aggregation • “misbehaving” flows: strong enough incentives? • ...or traffic shaping at input ports needed? hotnets2005_talk.fm Nov 14, 2005 12/14

Wrap Up FIFO Principle vs. Delay Control • ICDS reconciles both • incentives for traffic shaping, if low delay wanted • low-complexity QoS solution: single-class admission control Strong Game-theoretic Properties • with certain assumptions Background 2 Scheduler Model 4 Implementation 6 Implementation Details Evaluation 8 • partially solved Discussion 12 Wrap Up 13 Simulation Results • limited but encouraging hotnets2005_talk.fm Nov 14, 2005 13/14

Open Issues Validity of Game-theoretic Model • realistic assumptions? Implementation Details • non-trivial feedback loop • arrival rate → service rate • loss → sending rate Background 2 • feasible general configuration? Scheduler Model 4 • cf. Validity of Game-theoretic Model Implementation 6 • implementation efficiency Evaluation 8 • especially strict delay mode Discussion 12 Wrap Up 13 Multiplexing and Traffic Aggregation • robustness? hotnets2005_talk.fm Nov 14, 2005 14/14