CS-1652 Jack Lange University of Pittsburgh The slides are adapted from the publisher’s material All material copyright 1996-2009 J.F Kurose and K.W. Ross, All Rights Reserved

Principles of Congestion Control

Principles of Congestion Control informally: “too many sources sending too much data too fast for network to handle” different from flow control! manifestations: lost packets (buffer overflow at routers) long delays (queueing in router buffers) a top-10 problem! Transport Layer

Causes/costs of congestion: scenario 1 unlimited shared output link buffers Host A lin : original data Host B lout two senders, two receivers one router, infinite buffers no retransmission large delays when congested maximum achievable throughput Transport Layer

Causes/costs of congestion: scenario 2 one router, finite buffers sender retransmission of lost packet Network traffic includes retransmissions (l'in > lin) Host A lout lin : original data l'in : original data, plus retransmitted data Host B finite shared output link buffers Transport Layer

Causes/costs of congestion: scenario 2 lout lin idealization: perfect knowledge sender sends only when router buffers available lin : original data lout copy l'in: original data, plus retransmitted data A free buffer space! finite shared output link buffers Host B Transport Layer

Causes/costs of congestion: scenario 2 Idealization: known loss packets can be lost, dropped at router due to full buffers sender only resends if packet known to be lost lin : original data lout copy l'in: original data, plus retransmitted data A no buffer space! Host B Transport Layer

Causes/costs of congestion: scenario 2 Idealization: known loss packets can be lost, dropped at router due to full buffers sender only resends if packet known to be lost R/2 lin lout when sending at R/2, some packets are retransmissions but asymptotic goodput is still R/2 (why?) lin : original data lout l'in: original data, plus retransmitted data A free buffer space! Host B Transport Layer

Causes/costs of congestion: scenario 2 Realistic: duplicates packets can be lost, dropped at router due to full buffers sender times out prematurely, sending two copies, both of which are delivered R/2 when sending at R/2, some packets are retransmissions including duplicated that are delivered! lout lin R/2 timeout lin lout copy l'in A free buffer space! Host B Transport Layer

Causes/costs of congestion: scenario 2 Realistic: duplicates packets can be lost, dropped at router due to full buffers sender times out prematurely, sending two copies, both of which are delivered R/2 when sending at R/2, some packets are retransmissions including duplicated that are delivered! lout lin R/2 “costs” of congestion: more work (retrans) for given “goodput” unneeded retransmissions: link carries multiple copies of pkt decreasing goodput Transport Layer

Causes/costs of congestion: scenario 3 Q: what happens as lin and lin’increase? four senders multihop paths timeout/retransmit A: as red lin’ increases, all arriving blue pkts at upper queue are dropped, blue throughput g 0 Host A lout lin : original data Host B l'in: original data, plus retransmitted data finite shared output link buffers Host D Host C Transport Layer

Causes/costs of congestion: scenario 3 Host A lout Host B Another “cost” of congestion: when packet dropped, any “upstream transmission capacity used for that packet was wasted! Transport Layer

Approaches towards congestion control Two broad approaches towards congestion control: End-end congestion control: no explicit feedback from network congestion inferred from end-system observed loss, delay approach taken by TCP Network-assisted congestion control: routers provide feedback to end systems single bit indicating congestion (SNA, DECbit, TCP/IP ECN, ATM) explicit rate sender should send at Transport Layer

Explicit Congestion Notification (ECN) network-assisted congestion control: two bits in IP header (ToS field) marked by network router to indicate congestion congestion indication carried to receiving host receiver (seeing congestion indication in IP datagram) ) sets ECE bit on receiver-to-sender ACK segment to notify sender of congestion TCP ACK segment source destination application transport network link physical application transport network link physical ECE=1 ECN=11 ECN=11 ECN=00 IP datagram Transport Layer

TCP Congestion Control

TCP congestion control: additive increase multiplicative decrease approach: sender increases transmission rate (window size), probing for usable bandwidth, until loss occurs additive increase: increase CongWin by 1 MSS every RTT until loss detected multiplicative decrease: cut CongWin in half after loss additively increase window size … …. until loss occurs (then cut window in half) AIMD saw tooth behavior: probing for bandwidth congestion window size cwnd: TCP sender time Transport Layer

TCP Congestion Control: details sender limits transmission: Roughly: send CongWin bytes, wait RTT for ACKS, then send more bytes CongWin is dynamic, function of perceived network congestion How does sender perceive congestion? loss event = timeout or 3 duplicate acks TCP sender reduces rate (CongWin) after loss event three mechanisms: AIMD slow start conservative after timeout events LastByteSent - LastByteAcked  CongWin rate = CongWin RTT Bytes/sec Why is congwin not in the packet headers? Transport Layer

TCP Slow Start Increase rate exponentially fast until first loss event or CongWin reaches ss-thresh Loss 3 dup ACKs Timeout ssthresh – cwnd size at last congestion When connection begins, CongWin = 1 MSS Example: MSS = 500 bytes & RTT = 200 msec initial rate = 20 kbps available bandwidth may be >> MSS/RTT desirable to quickly ramp up to respectable rate (500 / .2 ) * 8 = 20kbps Why are timeouts and duplicate acks treated differently? Transport Layer

TCP Slow Start Host A Host B when connection begins, increase rate exponentially until first loss event: Initially: CongWin = 1 MSS every RTT: CongWin = 2 * CongWin done by incrementing CongWin for every ACK received summary: initial rate is slow but ramps up exponentially fast one segment RTT two segments four segments time Transport Layer

Refinement: inferring loss Philosophy: TCP RENO After 3 dup ACKs: CongWin = CongWin / 2 window then grows linearly But after timeout event: CongWin = 1 window then grows exponentially to a threshold, then grows linearly TCP TAHOE After 3 dup ACKs OR a timeout event: 3 dup ACKs indicates network capable of delivering some segments timeout indicates a “more alarming” congestion scenario Transport Layer

TCP: switching from slow start to CA Q: when should the exponential increase switch to linear? A: when cwnd gets to 1/2 of its value before timeout. Implementation: variable ssthresh on loss event, ssthresh is set to 1/2 of cwnd just before loss event Transport Layer

Summary: TCP Congestion Control New ACK! congestion avoidance cwnd = cwnd + MSS (MSS/cwnd) dupACKcount = 0 transmit new segment(s), as allowed new ACK . dupACKcount++ duplicate ACK slow start timeout ssthresh = cwnd/2 cwnd = 1 MSS dupACKcount = 0 retransmit missing segment cwnd = cwnd+MSS transmit new segment(s), as allowed new ACK dupACKcount++ duplicate ACK L ssthresh = 64 KB L cwnd > ssthresh timeout ssthresh = cwnd/2 cwnd = 1 MSS dupACKcount = 0 retransmit missing segment ssthresh= cwnd/2 cwnd = ssthresh + 3 retransmit missing segment dupACKcount == 3 timeout ssthresh = cwnd/2 cwnd = 1 dupACKcount = 0 cwnd = ssthresh dupACKcount = 0 New ACK ssthresh= cwnd/2 cwnd = ssthresh + 3 retransmit missing segment dupACKcount == 3 fast recovery cwnd = cwnd + MSS transmit new segment(s), as allowed duplicate ACK Transport Layer

Summary: TCP Congestion Control When CongWin is below Threshold, sender in slow- start phase, window grows exponentially. When CongWin is above Threshold, sender is in congestion-avoidance phase, window grows linearly. When a triple duplicate ACK occurs, Threshold set to CongWin/2 and CongWin set to Threshold. When timeout occurs, Threshold set to CongWin/2 and CongWin is set to 1 MSS. Transport Layer

TCP sender congestion control State Event TCP Sender Action Commentary Slow Start (SS) ACK receipt for previously unacked data CongWin = CongWin + MSS, If (CongWin > Threshold) set state to “Congestion Avoidance” Resulting in a doubling of CongWin every RTT Congestion Avoidance (CA) CongWin = CongWin+MSS * (MSS/CongWin) Additive increase, resulting in increase of CongWin by 1 MSS every RTT SS or CA Loss event detected by triple duplicate ACK Threshold = CongWin/2, CongWin = Threshold, Set state to “Congestion Avoidance” Fast recovery, implementing multiplicative decrease. CongWin will not drop below 1 MSS. Timeout CongWin = 1 MSS, Set state to “Slow Start” Enter slow start Duplicate ACK Increment duplicate ACK count for segment being acked CongWin and Threshold not changed Transport Layer

TCP throughput avg. TCP thruput as function of window size, RTT? ignore slow start, assume always data to send W: window size (measured in bytes) where loss occurs avg. window size (# in-flight bytes) is ¾ W avg. throughput is 3/4W per RTT avg TCP thruput = 3 4 W RTT bytes/sec W W/2 Transport Layer

TCP Futures: TCP over “long, fat pipes” example: 1500 byte segments, 100ms RTT, want 10 Gbps throughput requires W = 83,333 in-flight segments throughput in terms of segment loss probability, L [Mathis 1997]: ➜ to achieve 10 Gbps throughput, need a loss rate of L = 2·10-10 – a very small loss rate! new versions of TCP for high-speed TCP throughput = 1.22 . MSS RTT L Transport Layer

TCP Fairness fairness goal: if K TCP sessions share same bottleneck link of bandwidth R, each should have average rate of R/K TCP connection 1 bottleneck router capacity R TCP connection 2 Transport Layer

Why is TCP fair? Two competing sessions: Additive increase gives slope of 1, as throughout increases multiplicative decrease decreases throughput proportionally R equal bandwidth share loss: decrease window by factor of 2 congestion avoidance: additive increase Connection 2 throughput What is capacity? What is the equal share? loss: decrease window by factor of 2 congestion avoidance: additive increase Connection 1 throughput R Transport Layer

Fairness (more) Fairness and parallel TCP connections Fairness and UDP nothing prevents app from opening parallel connections between 2 hosts. Web browsers do this Example: link of rate R supporting 9 connections; new app asks for 1 TCP, gets rate R/10 new app asks for 11 TCPs, gets R/2 ! Fairness and UDP Multimedia apps often do not use TCP do not want rate throttled by congestion control Instead use UDP: pump audio/video at constant rate, tolerate packet loss Research area: TCP friendly Transport Layer

Chapter 3: Summary principles behind transport layer services: multiplexing, demultiplexing reliable data transfer flow control congestion control instantiation and implementation in the Internet UDP TCP Next: leaving the network “edge” (application, transport layers) into the network “core” Transport Layer