Wireless TCP KAIST CS 644 Advanced Topics in

Wireless TCP KAIST CS 644 Advanced Topics in Networking Jeonghoon Mo <jhmo@icu. ac. kr> School of Engineering Information and Communications University 1

Overview TCP Basics l Challenges in Wireless Networks High Bit Error Rate Variable Delays Mobility Interaction with other Layers l Jeonghoon Mo October 2004 2

TCP Basics l l l Transmission Control Protocol A Protocol for Reliable Data Transmission Two important roles Congestion Control for efficient and fair bandwidth sharing l l l Addtive increase and multiplicative decrease (AIMD) of window Fast recovery Fast retransmit Loss Recovery for Reliable Transmission l l l Jeonghoon Mo October 2004 Timer-based Retransmission RTO = mean (RTT) + 4 Std. Dev(RTT) Cumulative ACK from receiver 3

Challenges in Wireless Networks High Bit Error Rate l Variable Delays l Mobility (Handoff) l Interaction with Lower Layer Protocol (MAC) l Jeonghoon Mo October 2004 4

High Bit Error Rate Challenge Physical Property of Wireless Medium Raleigh fading, Slow Fading, Interference l TCP Throughput 1/sqrt(p) l Very poor performance with High Bit Error Rate l Jeonghoon Mo October 2004 5

Responses – High Bit Errors Extensive Research (599 results in IEEE Explorer) l Link Layer Retransmissions/ FEC l Differentiation of Wireless and Congestion Losses Explicit Notification End-to-End Differentiations l Use of Available Bandwidth Estimate l Jeonghoon Mo October 2004 6

Link Layer Retransmission/FEC Link Layer Solution Little or No Modifications in the TCP stack l Cellular Networks uses ARQ/FEC. l TCP Snoop (95), I-TCP l TCP Host BS Jeonghoon Mo October 2004 Link Layer Rxmt 7

Link Layer Retransmission l Link Layer Retransmission helps improving performance. 128 Retrx Limit 16 Retrx Limit Source: Wireless TCP Performance with Link Layer FEC/ARQ (ICC 1999) Jeonghoon Mo October 2004 8

Forward Error Correction Westwood+ performs better than TCP with FEC. l Performance Improvement comes more from good ABE algorithm than FEC. l Source: Adaptive End-to-End FEC for Improving TCP Performance over Wireless Links (ICC 2004) Jeonghoon Mo October 2004 9

Wireless Loss vs. Congestion Loss l IDEAL TCP If (congestion) Perform congestion recovery If (loss from ERROR) Perform loss recovery l CURRENT TCP (TCP Reno) If (congestion || loss from ERROR) Perform congestion recovery and loss recovery Jeonghoon Mo October 2004 10

How to Differentiate them? Explicit Notification l End-to-End Differentiation l Jeonghoon Mo October 2004 11

Explicit Notifications l l l Router or Access Point notifies TCP whether or not the loss is from Wireless Error. Motivated from the ECN proposal Usually, BS or AP needs some modifications Deployment Issue Next Generation Network will use feedback in the header. Examples Explicit Loss Notification (ELN) Explicit Bad State Notification (ESBN) XCP, TCP Jersey, FAST Jeonghoon Mo October 2004 12

XCP XCP: An e. Xplicit Control Protocol Explicit feedback 1. 2. Congestion Controller Fairness Controller (Source: XCP Presentation) Jeonghoon Mo October 2004 13

XCP Round Trip Round Time Trip Time Congestion Window Feedback = + 0. 1 packet Congestion Header Jeonghoon Mo October 2004 (Source: XCP Presentation) 14

XCP Round Trip Time Congestion Window Feedback == Feedback + 0. 3 0. 1 packet (Source: XCP Presentation) Jeonghoon Mo October 2004 15

XCP Congestion Window = Congestion Window + Feedback XCP extends ECN and CSFQ Routers compute feedback without any per-flow state (Source: XCP Presentation) Jeonghoon Mo October 2004 16

End-to-End Differentiation To overcome deployment issue of ECN, use only available information to the end hosts. l Sender or Receiver tries to differentiate two different losses based on information such as interarrival time of lost packets [biaz 99] use rtt, window size, loss pattern l Jeonghoon Mo October 2004 17

Some Ideas of E 2 E Differentiation l l Vegas: use “w-xd” to measure congestion Biaz: l Spike: l Jitter-based TCP (JTCP) w· (queue_ratio) > k then congestion loss queue_ratio = (interarrival – interdeparture)/ interdeparture Jeonghoon Mo October 2004 18

Jitter-based TCP (Wu 04) Similar to Vegas. l However, uses jitter ratio measurement with the help from the receiver l If the expected queue ratio is larger than the threshold, the loss is from the congestion. l Jeonghoon Mo October 2004 19

Summary ECN Deployment Issue Will be used in the Future l E 2 E Differentiation Has Limitations in differentiating congestion losses from the wireless losses. l Jeonghoon Mo October 2004 20

Available Bandwidth Estimation l AIMD -> AIAD adaptive decrease algorithm After 3 DUPs, ssthresh =ABE() * RTTmin cwin = min (cwin, ssthresh) - After TIMEOUT ssthresh =ABE() * RTTmin cwin = 1 l TCP Westwood(01), TCP Jersey Jeonghoon Mo October 2004 21

ABE Algorithm : arrival Time of the n-th ACK : acked amount by the n-th ACK : bandwidth estimate Jeonghoon Mo October 2004 22

TCP Westwood Accuracy of ABE Throughput Gain source: TCP Westwood: Congestion Window Control Using Bandwidth Estimation Jeonghoon Mo October 2004 23

TCP Jersey (04) l Similar to TCP Westwood in that it uses ABE technique Their estimate method is a bit different l Uses Congestion Warning simpler version of RED Jersey source: TCP-Jersey for Wireless IP Communications Jeonghoon Mo October 2004 24

Decoupled TCP (Mo, Kang & Kim 04) l Motivations How to differentiate has received attentions. What to do with the differentiation has not. Low performance of the newest protocols at high loss rate l Jeonghoon Mo October 2004 TCP Westwood gets about 10% of whole bandwidth when the loss rate is 10%. 25

Decoupled TCP l Two Ideas: Decouple Loss Recovery from Congestion Control (NETBLT) New ABE Algorithm Jeonghoon Mo October 2004 26

New ABE Algorithm of Westwood ABE Algorithm of DTCP Westwood Jeonghoon Mo October 2004 27

Decoupled TCP Reno, Sack DTCP Westwood Jeonghoon Mo October 2004 28

Challenges in Wireless Networks High Bit Error Rate l Variable Delays l Mobility (Handoff) l Interaction with Lower Layer Protocol (MAC) l Jeonghoon Mo October 2004 29

Variable Delays TCP sets RTO = mean(RTT) + 4 std. dev(RTT) l High delay variation causes spurious TIMEOUT. l l Spurious TIMEOUT causes unnecessary retransmission and rate reduction Jeonghoon Mo October 2004 30

TCP-Eifel: Variable Delays Eifel: A name of Mt. range in W. Germany l Use timestamp option The sender timestamps packets and stores times of retransmitted packets. The receiver echoes back the timestamp. l Jeonghoon Mo October 2004 31

Challenges in Wireless Networks High Bit Error Rate l Variable Delays l Mobility (Handoff) l Interaction with Lower Layer Protocol (MAC) l Jeonghoon Mo October 2004 32

Mobility As we move towards Ubiquitous paradigm, seamless service is considered more important. l Supporting mobile users with Qo. S has been an issue. l Impact of Mobility on TCP Performance Mobile IP Ad-hoc wireless Networks Hand-off l l Approaches to improve TCP performance M-TCP Freeze TCP Use of Multicast Jeonghoon Mo October 2004 33

Mobility Increase the Capacity of Ad -hoc Wireless Networks Best paper award at INFOCOM 2001 l Showed that throughput can be increase via mobile relay nodes. l Can TCP exploit this results? ? ? l (Grossglauser and Tse, 01) Jeonghoon Mo October 2004 34

TCP Throughput vs. Mobile Speed Decreasing function of speed Not Always source: Analysis of TCP Performance over Mobile Ad Hoc Networks Jeonghoon Mo October 2004 35

Why this happens? The TCP throughput decreases with movement because after route failure, it takes time for TCP to recover during which TCP fails to utilize bandwidth l Why not always, sometimes after reroute, the new path can be better than old one. Mobility increases the capacity of Ad-hoc Wireless networks [DTSE 2001] l Jeonghoon Mo October 2004 36

Freeze-TCP (Goff, 00) Freeze transmission during hand-off Fixed Host (Sender) Probe res ZWP Connection BS BS MH ZWA MH MH MH ZWA: Zero Window Advertisement ZWP : Zero Window Probing (source: http: //home. postech. ac. kr/~really 97/Freeze. ppt) Jeonghoon Mo October 2004 37

Challenges in Wireless Networks High Bit Error Rate l Variable Delays l Mobility (Handoff) l Interaction with Lower Layer Protocol (MAC) l Jeonghoon Mo October 2004 38

Interaction with MAC Downlink of GPRS networks : TDMA l Downlink of CDMA 2000 HDR : PF scheduling l TCP performs AIMD cogestion control for fairness and Efficiency. l Is the AIMD control required? l Jeonghoon Mo October 2004 39

TCP Throughput vs. # of Hops in a 802. 11 Network Not Always source: Analysis of TCP Performance over Mobile Ad Hoc Networks Jeonghoon Mo October 2004 40

Challenges in Wireless Transport Jeonghoon Mo October 2004 41

BACKUP 42

AIMD Congestion Control Packet Loss window Cd+B Time Loss based Congestion detection: Increase the window size until packet loss. Additive Increase and Multiplicative Decrease (AIMD) of Window Size Jeonghoon Mo October 2004 43

Fast Retransmit RTO = mean(RTT) + 4 * std. dev(RTT) l The RTO value is too long. l Instead of waiting for the timer expiration, after 3 duplicate packets, the sender retransmit packets. l l can cause spurious retransmission problem in ad-hoc networks Jeonghoon Mo October 2004 44

Fast Recovery Inflate congestion window temporarily to allow packet to be transmitted after a packet loss. cwnd = ssthresh + number of dupacks l Why increase cwnd ? the starting point of sliding window stops. the sender cannot send until the lost packet is recovered due to window restriction can cause link under utilization. l Jeonghoon Mo October 2004 45

Round Trip Time Measurement Adaptive Retransmission Time Out l Keep track of mean(RTT) and Deviation(RTT) l Retransmission Time Out (RTO) RTO = mean(RTT) + 4 Deviation(RTT) Jeonghoon Mo October 2004 46

TCP Implementations l TCP DCR [TAMU ] Delayed Congestion Response Similar to Snoop in that it uses link layer retransmission Jeonghoon Mo October 2004 47

TCP-PR (Persistent Reordering) l does not use DUP as the signal of packet loss Jeonghoon Mo October 2004 48