EP2200 Performance analysis of Communication networks. Topic 3 Congestion and rate control

Transcription

1 EP00 Performance analysis of Communication networks Toic 3 Congestion and rate control

2 Congestion, rate and error control Lecture material: Bertsekas, Gallager, Data networks, 6.- I. Kay, Stochastic modeling, J-Y Le Boudec, Rate adatation, congestion control and fairness: a tutorial, Nov. 005

3 Control functions in communication networks fairness concet congestion control rate control admission control error control delay control Differentiated transm. (congestion control) (error control) (admission control) 3

4 Congestion control To regulate the acket oulation in the network To share resources (link bandwidth, buffer sace) Flow control: between two users for seed matching, sometimes means congestion control in the literature The main objectives of congestion control Efficiency high utilization (from the network rovider s ersective) high er flow throughut, low delay (from users ersective) Fairness: fair allocation of resources 4

5 Congestion control Throughut Queues building u Congestion collase Offered load What haens if the incoming traffic is not restricted? Bottleneck links: the offered traffic is higher than the link transmission caacity: temorarily (bursts arriving) or ermanently What haens at bottleneck links? queue sizes grow, end-to-end delays increase queue sace fills u, ackets get droed ackets are retransmitted by the alications, further increasing the load Congestion collase: the network throughut decreases and delays become excessive 5

6 Congestion control techniques Should deend on the service requirements of the alication voice, video (streaming): minimum bandwidth requirement, some loss is allowed, delay sensitive data (elastic): lossless end-to-end transmission required Call blocking at the network edge + rate control calls blocked if resources are not available the rate of acceted calls is controlled Packet discarding at a network node at buffer overflow or earlier discarding olicy - fairness, service differentiation Packet blocking at the network edge acket waits in a queue outside the network Grou work: select solutions for streaming and for elastic flows find examles for the techniques above 6

7 Congestion control techniques Should deend on the service requirements of the alication voice, video: minimum bandwidth requirement, delay sensitive data: requires strict error control Call blocking at the network edge + rate control calls blocked if resources are not available the rate of acceted calls is controlled Packet discarding at a network node buffer overflow or active queue management discarding olicy - fairness, service differentiation Packet blocking at the network edge acket waits in a queue outside the network 8

8 End-to-end window based congestion control A Data ACK B Basic idea: Window: Uer bound on the number of ackets transmitted by A and not yet acknowledged by B Inut rate reduced if acknowledgements arrive slowly achieves congestion control 9

9 Fixed window based congestion control Congestion is indicated by increased round-tri time (RTT) Simle case: ACK after each acket recetion Window size W is * decremented after each transmission * incremented if ACK arrives Transmission, if W > 0 No lost Data acket or lost ACK Parameters constant maximum window size (W m ) constant acket transmission time (: acket size/link rate [sec]) round-tri time (rtt: roagation, transmission and queuing delays) To calculate: maximum acket transmission rate (r: [acket/sec]) r min Wm, rtt 0

10 Fixed window based congestion control Maximum acket transmission rate (r: [acket/sec]) Parameters constant maximum window size (W m ) constant acket transmission time (: acket size/link rate [sec]) round-tri delay (rtt: roagation, transmission and queuing delays) r min Wm, rtt Rate inversely roortional to round-tri delay Congestion control: large queuing delays large rtt lower rate Large W m allows higher rate if there is no congestion Reacts to congestion in W m ackets transmission time large window means slow reaction to congestion What rtt value means congestion? When should the congestion control be activated? (rttw m )?

11 Window control - acket delay End-to-end acket delay T? Delay: roagation, transmission and waiting at the queues in the multiho ath Queuing analysis in steady state, for given network load Model the network as a black box, aly Little theorem (N T) Assumtion: Constant arrival intensity (er node rate) Infinite buffers, no loss N T n i N λ const W n i n: active flows i i const W λ W i : max window size for flow i N: number of ackets in the network λ: aggregate throughut of all flows (no loss infinite buffers!) i i For constant aggregate throughut (λ) Delay is roortional to the number of active sessions and to the window sizes Large W m - large delay

12 Window control How to select the maximum window size? Small window: OK acket delay, but low throughut (congestion control starts with low rtt) Large window: OK throughut, but high acket delay and long reaction time r min, T n i const i W i λ W rtt Dynamic maximum window sizes are necessary to follow the network load small window if the network is congested large window if the network is low loaded congestion is controlled by the window size not by the rtt But what indicates the congestion: acket dro, increased rtt,? 3

13 Congestion control-dynamic window size How to select the window size? Small window: OK acket delay, but low throughut Large window: OK throughut, but high acket delay Dynamic window sizes are necessary to follow the network load Congestion is controlled by the window size not by the rtt TCP Performance with dynamic window sizes - comes now More realistic models for TCP erformance reading assignment 4

14 TCP congestion control Window based congestion control Dynamic window size: decrease max window size if congestion is detected congestion indicated by acket loss increase max window size if current rate does not cause congestion (e.g., no loss) How to increase and decrease the window sizes? Additive-increase, multilicative decrease (AIMD) Efficient and fair (if all users get the same immediate indication of congestion): Chiu and Jain, 989 Probing - increase hase in each rtt: Congestion - decrease hase: TCP: a0.5, b w i w + + b,0 < b << w i max w i awi,0 < + a < TCP additional hases: slow start, fast recovery not considered now 5

15 Analysis AIDM model Congestion indicator: acket loss (full buffer on the ath) Question: How does the throughut deend on the loss rate? Assumtions for a very simlified case saturated source (always has acket to send) constant background traffic (all other traffic), loss at the same window size loss due to congestion only loss is the only congestion indicator transients negligible (long flows) we model a static congestion avoidance hase constant round tri time (rtt) rtt >> transmission time constant acket size L low loss robability (to simlify calculations) W m W m / W increased to W+ after rtt if there was no loss W decreased to W/ after rtt if there was a loss One acket loss here W ackets transmitted within an rtt t 6

16 7 TCP throughut vs. acket loss Th as a function of and rtt rtt L rtt L rtt L T L N Th N W W W W N rtt W T N N T L N Th m m m m m ) ( , ) ( rtt: round tri time (s) : acket loss robability L: acket length (bit) Th: throughut (bit/s) W m : max achieved window (ackets) T 0 : cycle (s) time between losses N: ackets transmitted in one cycle w(t) t W m W m / T 0

17 TCP throughut vs. acket loss Statement: throughut in steady state varies as the inverse-square-root of the loss rate, and is inversely roortional to the round tri time (this we have seen for fixed window already!) Th 3 L rtt

18 Recall and outline for today Last lecture: Congestion control definition and ossible solutions End-to-end congestion control based on acket blocking at the network edge Fixed window size schemes throughut, delay conflict Dynamic window congestion control Simle AIMD scheme and throughut analysis first stes towards evaluating TCP Todays lecture: Intro for the TCP modeling home reading TCP friendly rate control for delay sensitive traffic Rate control for guaranteed service Home reading: J. Padhye, F. Firoiu, D. Towsley, J. Kurose, "Modeling TCP throughut: a simle model and its emirical validation," Sigcomm, 998 9

19 TCP friendly rate control Some flows can not allow acket blocking at the network edge E.g., streaming voice and video Without congestion control they could monoolize the network by ressing back TCP flows Idea: Rate control: in average the sending rate should be the same as for a TCP flow in the same situation (rtt, loss rate) The alication adats coding rate accordingly Th 3 L rtt

20 Project toics related to TCP Research areas where some of the assumtions are released or discussed: TCP erformance in high bandwidth-delay roduct networks (slow feedback) loss does not mean congestion: TCP over wireless links TCP for short flows (never leave the slow start hase) TCP and self similarity loss distribution among sessions: active queue management for TCP 6

21 Control functions in communication networks fairness concet congestion control rate control admission control error control delay control Differentiated transm. (congestion control) (error control) (admission control) 7

22 Rate control and guaranteed service Contract between the user and the network If the user satisfies a traffic constraint, Then the network rovides delay, delay jitter (and loss) limits Rate control during the connection time Peak rate Average rate Burstiness (e.g., max. number of ackets transmitted without time ga) Window based control (r,t) Rate r ensured in window T (n maximum number of ackets er T: nrt) Juming or sliding window versions Problem: Burstiness: juming: bnrt, sliding: bnrt Max burst size can not be controlled indeendently from r and T Traffic enveloe max transmitted data within any time interval is given by a function b(t) how to imlement it? 8

23 Rate control Leaky-bucket Leaky-bucket scheme Version : average rate and max burst size token buffer size (b) token generation rate (r) ackets from inut enter data buffer, waiting for service acket transmitted to outut if there is a token in the token buffer token removed at acket transmission Result: traffic enveloe: b(t)b+rt average rate: (b+rt)/t Note: from outut the ackets enter a transmission buffer (MAC) Note: t is not the time from the system start, but any interval t during the transmission 9

24 Rate control Leaky-bucket Leaky-bucket scheme Version : eak rate, average rate, burstiness average rate buffer (b) average token generation rate (r) eak rate buffer () eak token generation rate (P>r) ackets from inut enter data buffer, waiting for service acket transmitted to outut if there is at least one token in both token buffers tokens removed at acket transmission Result: traffic enveloe b(t) 30

25 Summary Congestion control to limit the traffic in the network (avoid loss or large delays) window based congestion control AIMD schemes and TCP Rate control to control the average rate, burstiness or eak rate of traffic injected to the network with the goal of roviding service guarantees Solutions window based roblem with rate and burstiness couling traffic enveloe based leaky bucket At home J. Padhye, F. Firoiu, D. Towsley, J. Kurose, "Modeling TCP throughut: a simle model and its emirical validation," Sigcomm, 998 Check the matlab emulator for the Leaky Bucket! 3