Reliable Transport II: TCP and Congestion Control

Transcription

1 Reliable Transport II: TCP and Congestion Control Stefano Vissicchio UCL Computer Science COMP0023

2 Recap: Last Lecture Transport Concepts Layering context Transport goals Transport mechanisms and design choices TCP Protocol (part I) Packet header format Connection establishment and teardown 2

3 Recap: TCP s Many End-to-End Goals Provide a connection-oriented service to applications Not totally insecure.. Ensure Reliability Recover from data loss Avoid receipt of duplicated data Preserve data ordering Provide integrity against corruption Transfer data as fast as possible Avoid sending faster than receiver can accept data Avoid congesting network 3

4 Recap: TCP Packet Header TCP header: 20 bytes long (minimum) Encapsulate data, encapsulated into an IP header Checksum covers header + pseudo header 4

5 TCP: Data Transmission Each byte numbered sequentially, mod 2 32 Sequence number (seqno) field in TCP header indicates first user payload byte in packet Both sender and receiver buffer data Sender buffers data in case retransmission s required Receiver buffers data for in-order reassembly Buffer space à sender s and receiver s window size Receiver indicates receive window sizes explicitly in window field in TCP header 5

6 TCP: Data Transmission (cont d) Sender uses a sliding window that does not exceed send and receive window sizes.. and network s estimated capacity (details later today) window s low end advances as packets are ACKed, high end advances as receive window increases Receiver sends cumulative ACKs ACK number in TCP header names highest contiguous byte number received thus far +1 one ACK per received packet, or delayed ACK delayed ACK: receiver batches ACKs, sends one for every pair of data packets (200 ms max delay) 6

7 TCP: Data Transmission (cont d) Sender uses a sliding window that does not exceed send and receive window sizes.. and network s estimated capacity (details later today) window s low end advances as packets are ACKed, high end advances as receive window increases So how fast to transmit? Receiver sends cumulative ACKs ACK number in TCP header names highest contiguous byte number received thus far +1 one ACK per received packet, or delayed ACK delayed ACK: receiver batches ACKs, sends one for every pair of data packets (200 ms max delay) 7

8 Outline Slow Start and Retransmit Timeouts AIMD Congestion control Throughput, loss, and RTT equation Protocol state machine 8

9 Undesirable Retransmit Behavior Original TCP, before [Jacobson 88]: at start of connection, send full window of packets retransmit each packet immediately after its timer expires with the Go-Back-N strategy Result: window-sized bursts of packets sent into network 9

10 Sequence no. (KB) Pre-Jacobson TCP with Single Sender (Obsolete!) Time-sequence plot taken at sender Orange line: available 20 KB/s capacity Time (s) 10

11 Sequence no. (KB) Pre-Jacobson TCP with Single Sender (Obsolete!) Bursts of packets Spurious retransmits Same packets sent again and again Time (s) 11

12 Sequence no. (KB) Pre-Jacobson TCP with Single Sender (Obsolete!) Two main root causes: Retransmit timeouts Transmission Rate Bursts of packets Spurious retransmits Same packets sent again and again Time (s) 12

13 TCP: Retransmit Timeouts Sender sets timer for each sent packet if timer expires before ACK returns, packet resent Expected time for ACK to return: RTT TCP estimates RTT using EWMA: measurements m i from timed packet/ack pairs estimated RTT: RTT i = (1-α) RTT i-1 + α m i with α = 1/8 Original TCP set retransmit timeout (RTO) RTO i = β RTT i, with β = 2 Is this accurate enough? Recall dangers of too-short and too-long RTT estimates 13

14 Mean and Variance: Jacobson s Findings Above link load of 30% at router, β RTT i will retransmit too early Response to increasing load: waste bandwidth on duplicate packets Result: congestion collapse!! [Jacobson 88]: estimate v i, mean deviation (EWMA of m i RTT i ), stand-in for variance v i = v i-1 (1 - γ) + γ m i -RTT i Use RTO i = RTT i + 4v i 14

15 Mean and Variance: Jacobson s RTT Estimator Above link load of 30% at router, β RTT i will retransmit too early Response to increasing load: waste bandwidth on duplicate packets Result: congestion collapse!! Main solution [Jacobson 88]: Estimate v i = mean deviation as EWMA of m i RTT i, to approximate RTT variance v i = v i-1 (1 - γ) + γ m i -RTT i Use RTO i = RTT i + 4v i 15

16 Mean and Variance: Jacobson s RTT Estimator Above link load of 30% at router, β RTT i will retransmit too early Mean and Variance RTT estimator Response to increasing load: waste bandwidth on duplicate packets used by all modern TCPs Result: congestion collapse!! Main solution [Jacobson 88]: Estimate v i = mean deviation as EWMA of m i RTT i, to approximate RTT variance v i = v i-1 (1 - γ) + γ m i -RTT i Use RTO i = RTT i + 4v i 16

17 Sequence no. (KB) Pre-Jacobson TCP with Single Sender (Obsolete!) Two main root causes: Retransmit timeouts Transmission Rate Bursts of packets Spurious retransmits Same packets sent again and again Time (s) 17

18 Self-Clocking: Conservation of Packets Goal: self-clocking transmission as each ACK returns, one data packet is sent spacing of returning ACKs: matches spacing of packets in time at slowest link on path 18

19 Reaching Equilibrium: Slow Start At connection start, sender sets congestion window size, cwnd, to pktsize (one packet s worth of bytes) Sender sends up to minimum of receiver s advertised window size rwnd and cwnd Upon return of each ACK until receiver s advertised window size reached, increase cwnd by pktsize bytes Slow means exponential window increase! Takes log 2 (rwnd / pktsize) RTTs to reach receiver s advertised window size W If the bottleneck is the receiver... 19

20 Post-Jacobson TCP with Single Sender: Slow Start and Mean+Variance RTT Estimator Time-sequence plot at sender Slower start, but no spurious retransmits 20

21 Outline Slow Start and Retransmit Timeouts AIMD Congestion control Throughput, loss, and RTT equation Protocol state machine 21

22 Goals in Congestion Control High utilization Achieve high utilization: don t waste capacity! Fairness Divide bottleneck link capacity fairly among users Be stable: achieve steady allocation among users Avoid congestion collapse!! 22

23 Goals in Congestion Control High utilization Achieve high utilization: don t waste capacity! Fairness Divide bottleneck link capacity fairly among users Be stable: achieve steady allocation among users Avoid congestion collapse!!... With no information about bottleneck link(s), network state, and other users! 23

24 Congestion Collapse Saturation Knee Throughput (bps) Congestion collapse! Offered load (bps) Cliff behavior observed in [Jacobson 88] 24

25 Congestion Requires Slowing Senders Bigger buffers cannot prevent congestion Senders must slow to alleviate congestion Absence of ACKs implicitly indicates congestion TCP congestion window size determines sending rate Recall: correct window size is bottleneck bandwidthdelay product (or our share of it) How can sender learn correct window size? Search for it, by adapting window size Feedback from network: ACKs return (window OK) ACKs do not return (window too big) 25

26 Congestion Requires Slowing Senders Bigger buffers cannot prevent congestion Senders must slow to alleviate congestion Absence of ACKs implicitly indicates congestion TCP congestion window size determines sending rate Recall: correct window size is bottleneck bandwidthdelay product (or our share of it) How can a sender learn correct window size? Search for it, by adapting window size Feedback from network: ACKs return (window OK) ACKs do not return (window too big) 26

27 TCP s cwnd limits the number of packets TCP has sent for which it has not yet seen an acknowledgment. Data packets Acks for data packets

28 Avoiding Congestion: Multiplicative Decrease Sender uses sending window of size min(cwnd, rwnd), where rwnd is receiver s advertised window Upon timeout for sent packet, sender presumes the packet was lost to congestion, and: sets ssthresh = cwnd / 2 sets cwnd = pktsize uses slow start to grow cwnd back up to ssthresh End result: cwnd = cwnd / 2, via slow start Sender sends ~ one window per RTT; halving cwnd halves transmit rate 28

29 Avoiding Congestion: Additive Increase Drops indicate that TCP is sending more than its fair share of bottleneck There s no feedback to indicate TCP using less than its fair share of bottleneck Solution: speculatively increase cwnd as ACKs return Additive increase: for each returning ACK, cwnd = cwnd + (pktsize pktsize) / cwnd You get cwnd/pktsize acks per RTT, so this increases cwnd by ~pktsize bytes each RTT 29

30 Avoiding Congestion: Additive Increase Drops indicate that TCP is sending more than its fair share of bottleneck There s no feedback to indicate TCP using less than its fair share of bottleneck Solution: Combined speculatively algorithm: increase cwnd as ACKs return Additive Increase, increase: Multiplicative for each returning Decrease ACK, (AIMD) cwnd = cwnd + (pktsize pktsize) / cwnd You get cwnd/pktsize acks per RTT, so this increases cwnd by ~pktsize bytes each RTT 30

31 Refinement: Fast Retransmit (I) Sender must wait well over one RTT for timer to expire before loss detected TCP s minimum retransmit timeout: 1 second Another loss indication: duplicate ACKs Suppose sender sends 1, 2, 3, 4, 5, but 2 lost Receiver receives 1, 3, 4, 5 Receiver sends cumulative ACKs 2, 2, 2, 2 Refinement: avoid slow start if 3 duplicate ACKs Directly halve cwnd 31

32 Fast Retransmit (II) Upon arrival of 3 duplicate ACKs, sender: sets cwnd = cwnd / 2, with no slow start retransmits missing packet enters additive increase A data, seqno = 1 data, seqno = 513 data, seqno = 1025 data, seqno = 1537 data, seqno = 2049 ACK = 513 ACK = 513 ACK = 513 ACK = 513 data, seqno = 513 B time 32

33 AIMD in Action packets in flight 3 dup ACKs received multiplicative decrease window full (too many losses for fast RTX) slow start additive increase timeout back in slow start Sender searches for correct window size, constantly over time time 33

34 AIMD in Action: Competing Flows

35 Why AIMD? Other control rules are possible: E.g., MIMD, AIAD, Recall goals: Links fully utilized (efficient) Users share resources fairly Each TCP flow adapts its own window size independently Must choose a control that will always converge to an efficient and fair allocation of windows 35

36 Chiu-Jain Phase Plots Flow 1 Consider two flows sharing a bottleneck link Plot cwnd of each (approx. bandwidth) Efficiency: sum of two flows rates fixed Fairness: two flows rates equal Equi-Fairness: ratio of two flows rates fixed Flow 2 cwnd Flow 2 Equi-Fairness Line (MI) (AI) Underload Fairness Line Overload Optimum Efficiency Line Flow 1 cwnd 36

37 Chiu Jain: AIMD Flow 1 Flow 2 cwnd Fairness Line Flow 2 Efficiency Line Queue Overflow Line Flow 1 cwnd AIMD converges to optimum efficiency and fairness the plot assumes both flows lose packets during congestion 37

38 Chiu Jain: AIAD Flow 1 Flow 2 Fairness Line Efficiency Line Queue Overflow Line AIAD doesn t converge to optimum point! Similar oscillations for MIMD 38

39 Simulation of Two TCP Flows (if only one flow loses packets during congestion)

40 Result: Over time, TCP equalizes the windows of competing flows Flow 2 Flow 1 Window of flow 2 The two windows are similar most of the time Window of flow 1

41 Outline Slow Start and Retransmit Timeouts AIMD Congestion control Throughput, loss, and RTT equation Protocol state machine 41

42 Modeling Throughput, Loss, and RTT How do packet loss rate and RTT affect throughput TCP achieves? Assume: only fast retransmits so no timeouts (i.e., no slow starts in steady-state) 42

43 Steady State Evolution of Window Over Time W W/2 time Mean window size: 3W / 4 One window sent per RTT Mean throughput, T: 3W / 4 packets per RTT 3BW / 4 bytes per RTT (B = packet size in bytes) (3BW / 4) / RTT bytes per second W depends on loss rate T = 3B.W 4RTT 43

44 Loss and Window Size Assume no delayed ACKs, fixed RTT cwnd grows by one packet per RTT So it takes W / 2 RTTs to go from window size W / 2 to window size W; this period is one cycle How many packets sent in total in a cycle? ((3W / 4) / RTT) x (W / 2 x RTT) = 3W 2 /8 One loss per cycle (as window reaches W) p = 8 3W 2 loss rate: so: W = 8 3p 44

45 Throughput, Loss, and RTT Model Recall: W = 8 3p = p Throughput T = 3B.W 4RTT Substituting for W gives: T = 3 / 2B RTT p Consequences: Increased loss quickly reduces throughput At same bottleneck, flow with longer RTT achieves less throughput than flow with shorter RTT! 45

46 Outline Slow Start and Retransmit Timeouts AIMD Congestion control Throughput, loss, and RTT equation Protocol state machine 46

47 TCP: Protocol State Machine [RFC793] 47

48 Summary: TCP and Congestion Control Connection establishment and teardown Robustness against delayed packets crucial Round-trip time estimation EWMAs estimate both RTT mean and deviation Congestion detection at sender Timeout: retransmit timer expires, halve window, slow start from one packet Fast Retransmit: three duplicate ACKs, halve window, no slow start Search for optimal sending window size Additive increase, multiplicative decrease (AIMD) AIMD converges to high utilization, fair sharing 48

49 Reading: Read It! Jacobson V., Congestion Avoidance and Control A true classic in networking one of the most influential ideas in the area Paper is available in PDF on calendar page of the class web site