TCP for wireless links

Dario Maggiorini

dario@dico.unimi.it

TCP

TCP congestion control

The algorithm for TCP congestion control is the main reason we can use the Internet successfully today despite largely unpredictable user access patterns and despite resource bottlenecks and limitations. Without TCP congestion control, the Internet could have become history a long time ago.

Resource Management Solutions

• Handling congestion– pre-allocate resources so as to avoid congestion – control congestion if (and when) is occurs

• Two points of implementation– routers inside the network (queuing discipline) – hosts at the edges of the network (transport protocol)

Destination1.5-Mbps T1 link

Router

Source2

Source1

100-Mbps FDDI

10-Mbps Ethernet

Detecting Congestion

• Packet drops indicate congestion– Is that really true?– Why does it work?

Src Dst

Packet

Ack

Drop

Timeout! No Ack =

Congestion!

Controlling Congestion – The Effect of Window Size

• Note that sender’s window is equal to the number of sender packets in flight (in the network). Why?

A Window’s worth of packets

X acks

Window

X more packets

Source Destination

Controlling Congestion

• Reduce window less packets in the network

• Increase window more packets in the network

• Idea: Concept of a congestion window – window is smaller when congestion is larger and vice versa

Additive Increase, Multiplicative Decrease

• Each time a packet drop occurs, slash window size in half (multiplicative decrease)– Multiplicative decrease is necessary to avoid

congestion

• When no losses are observed, gradually increase window size (additive increase)

AIMD (cont)

• In practice: increment a little for each ACKIncrement = (MSS*MSS)/CongestionWindow

CongestionWindow += Increment

Source Destination

…

• Algorithm– increment CongestionWindow by

one packet per RTT (linear increase)– divide CongestionWindow by two

whenever a timeout occurs (multiplicative decrease)

TCP Reno (Jacobson 1990)

SStime

window

CA

SS: Slow StartCA: Congestion Avoidance Fast retransmission/fast recovery

Problems

• What should the window size be– Initially?– Upon packet loss and timeout?

• Pessimistic window size? (e.g., 1)– Additive increase is too slow in ramping up window

size – short connections will not fully utilize available bandwidth

• Optimistic window size?– Large initial burst may cause router queue overflow

Slow Start

• Objective: determine the available capacity quickly

• Idea:– Use CongestionThreshold as an

optimistic CongestionWindow estimate– begin with CongestionWindow = 1

packet– double CongestionWindow each RTT

(increment by 1 packet for each ACK)– When CongestionThreshold is

crossed, use additive increase

Source Destination

…

Fast Retransmit

• Problem: coarse-grain TCP timeouts lead to idle periods

• Fast retransmit: – Send an ack on every packet reception– Send duplicate of last ack when a packet

is received out of order– Use duplicate ACKs to trigger

retransmission

Packet 1

Packet 2

Packet 3

Packet 4

Packet 5

Packet 6

Retransmitpacket 3

ACK 1

ACK 2

ACK 2

ACK 2

ACK 6

ACK 2

Sender Receiver

Self Clocking and Slow Start

• Each packet’s transmission is “clocked” by an ACK – no bursts develop

W=1 W=2 W=4 W=5 W=6 W=7

Slow Start

…

Self Clocking in Operation

• Each packet’s transmission is “clocked” by an ACK – no bursts develop

W=32

……

Self Clocking Interrupted

• During timeouts, ACKs are drained from the network. Self clocking is interrupted. Next transmission causes a burst. Hence, slow start!

W=32

……Lost

Timeout

Retransmission

Ack32

16 packet burst

Cut window in 1/2

ACKsDrained!!

Self Clocking and Fast Retransmit / Fast Recovery

• When fast retransmit is used, the packet is retransmitted before all ACKs are drained. Slow start is not needed

W=32

……Lost

FastRetransmission

Cut window in 1/2

Problems with TCP Reno

• During slow start– Underutilizes and then swamps path

• No “right rate”: cwnd traces a sawtooth– Underutilizes path– Increases queuing delay– Causes loss, reducing throughput– Inherently biased against long RTTs

TCP Vegas

• Uses congestion avoidance instead of congestion control– Vegas: Congestion avoidance: Predict and

avoid congestion before it occurs– Tahoe, Reno: Congestion control: React to

congestion after it occurs

• Question: How to predict congestion?

TCP Vegas

• Idea: source watches for some sign that router’s queue is building up and congestion will happen too; e.g.,– RTT grows– sending rate

flattens

60

20

0.5 1.0 1.5 4.0 4.5 6.5 8.0

KB

Time (seconds)

Time (seconds)

70

304050

10

2.0 2.5 3.0 3.5 5.0 5.5 6.0 7.0 7.5 8.5

900

300100

0.5 1.0 1.5 4.0 4.5 6.5 8.0

Sen

din

g K

Bp

s 1100

500700

2.0 2.5 3.0 3.5 5.0 5.5 6.0 7.0 7.5 8.5

Time (seconds)0.5 1.0 1.5 4.0 4.5 6.5 8.0Q

ueu

e s

ize

in r

oute

r

5

10

2.0 2.5 3.0 3.5 5.0 5.5 6.0 7.0 7.5 8.5

TCP Vegas

• In congestion avoidance– cwnd = (actual rate)x(baseRTT) + 2 pkts– Each RTT, tweak cwnd by 1 pkt if needed

• During slow start– To reduce overshoot, increase cwnd only

every other RTT– Exit slow start when

• cwnd > (actual rate)x(baseRTT) + 1 pkt

TCP Vegas (Brakmo & Peterson 1994)

SStime

window

CA

• Converges, no retransmission• … provided buffer is large enough

TCP Westwood

• Sender side only modification of TCP Reno Congestion control that exploits end to end bandwidth estimation.

• The bandwidth is estimated by low pass filtering the rate of returning acks.

• The bandwidth is used to compute congestion window and slow start threshold.

TCP Westwood Overview

• Slow Start and Congestion window aware of Bandwidth at time of congestion

• The increase after congestion is additive but decrease Adaptive (AIAD) as compared to AIMD (Additive Increase Multiplicative decrease) of Reno

TCPW implementation• Sender side Bandwidth Estimation by measuring and low pass

filtering the rate of returning acks

• When 3 DUPACKS are receivedssthresh=(Bandwidth*RTT)/seg_size cwnd =ssthresh

• When a coarse timeout expires ssthresh =(B*RTT)/seg_size cwnd =1

• When acks are successfully received TCPW increases cwnd according to Reno’s congestion control algorithm

TCPW Advantage over Reno

• In case of sudden increase in bottleneck load, reduction can be more drastic then a reduction by half and can be less drastic in other cases. This features improves stability and utilization

Go Back N Protocol (GBN)1 2 3 4 1 2 3

1 2 3 4

1 2 3 4 1 2 3

2 3 4

New Packet

Time Out Copies (Go Back)Sender

Sender

Receiver

Receiver

Selective Repeat Protocol

1 2 3 4 5 2 6 7

1 23 4 52, 3, 4, 51

Time OutSender

Receiver

Wireless networks

Hidden node problem

• A and C can send to B but can’t hear each other

• CSMA will be ineffective here

BA C

A’s transmit range C’s transmit range

Exposed node problem

• B and C can hear each other but could safely send simultaneously to A, D

• Compare to spatial reuse

in cellular

BA C

B’s transmit range

C’s transmit range

D

StrategyCSMA with Collision Avoidance

Fundamentally, a less greedy approach than Ethernet

• When medium busy, choose random backoff interval

– Wait for that many idle timeslots to pass before sending

– Remember p-persistence … a refinement

CSMA/CA

• When a collision is inferred, retransmit with binary exponential backoff (like Ethernet)

– Use CRC and ACK from receiver to infer “no collision”

– Again, exponential backoff helps us adapt “p” as needed

TCP over wireless

The wireless problem

In Wireless lossy links, the sporadic losses are not due to congestion thus it leads to unnecessary window and transmission rate reduction

Sources of Errors in Wireless Links

• Pauses due to handoff between cells

• Packet losses due to futile transmissions: mobile host out of reach of other transceivers (little or no overlap between cells);

• Packet losses due to transmission errors in wireless links.

Improving the Performance of TCP

MSS 1 MSS 2

SH

MH

Cell 1 Cell 2

Smooth Handoff

• Cellular networks should strive to provide smooth handoffs in order to eliminate packet losses during cell crossings.

• No overlaps are also good!!!– High aggregate bandwidth: adjacent cells can

use the same portion of the spectrum;– Support for low-powered mobile receivers;– Accurate location information

Comparision of Mechanisms

• End-to-end protocols

• Split-connection protocols

• Link-layer protocols

• Hybrid protocols

End-to-end Protocols

• Sender is aware of the existence of wireless hops.

• Selective Acknowledgments (SACKs): sender can recover from multiple packet losses without resorting to a coarse timeout.

• Explicit Loss Notification (ELFN): the sender can distinguish between congestion and other forms of losses.

Split-connection Protocols

• Aims to hide any non-congestion-related losses from the TCP sender.

• TCP connection is split between a sender and receiver into two separate connections at the base station:– TCP connection over wired link;– Specialized protocol over wireless link.

I-TCP: Indirect TCP

MH

MSR

FH

• MH = Mobile Host• MSR = Mobile Support Router• FH = Fixed Host

I-TCP TCP

TCP/IP in Mobile Environment

• Main reason for throughput degradation:– Loss of TCP segments during cell crossovers,

especially with non-overlapped cells.

• Effects:– Lost segments trigger exponential back off

and congestion control at the transmitting host.

– Congestion recovery phase may last for several seconds.

Indirect Protocol

• Different flow control and congestion control for wireless and wired links;

• Separate transport protocol supports disconnections, moves and other wireless related features;

• MSR manages much of the overhead;

• Faster reaction to mobility due to proximity between MSR and MH.

I-TCP Basics

move

MSR-2

FH

MH

MH socket

MH

MH socket

MSR-1 MSR-1

MSR1mhsocket

MSR1fhsocket

MSR2fhsocket

MSR2mhsocket

FH socket

I-TCP Handoff

Regular TCP

Wireless TCP

Link-layer Protocols

• Two main classes:– Error correction using techniques such as

Forward Error Correction;– Retransmission of lost packets in response to

automatic repeat request (ARQ) messages.

• Tuned to the characteristics of the wireless link.

Hybrid Protocols: The Snoop Prootocol

• An agent monitors every packet and maintains a cache of TCP segments that have not yet been acknowledged.

• Packet loss is detected by the arrival of a small number of duplicate acks or by a local timeout.

• The agent retransmits the lost packet and suppresses the duplicate acks.

Observations

• TCP-aware link-layer protocol with selective acknowledgments performs the best;

• Split-connection approaches is not a requirement for good performance.

• Selective acknowledgment is very useful in lossy links, especially for burst losses.

• Explicit Loss Notification is worth to try.

TCP over satellite links

Why Satellite?

• Affordable Access to Interactive Broadband Communication to all Areas of the Earth

• To Serve users who can not get Service economically by other means

• Satellite Types

• Transmission Quality

Altitude (km) RTT (ms)

GEO 36,000 588

MEO 10,390 250

LEO 1,375 70

Characteristics of Satellite Network

• Long Feedback Loop

• Large Delay-Bandwidth Product

• Transmission Errors

• Variable RTT

• Frequent Hand off

• Asymmetric Use

Corresponding Problems

• Slow Start– For single TCP– For HTTP applications

• Congestion Avoidance– Congestion vs. Transmission Error

• TCP Window Size

TCP Window Size

• Throughput = Window size / RTT• Max TCP Window Size = 216 = 65535 (bytes)

• Throughput <= 65535 bytes / 560 ms

= 117,027 bytes /second

Cannot utilize a T1 rate GEO satellite

Standard TCP Enhancements

• Fast Retransmit & Fast Recovery

• TCP SACKs

• TCP Large Windows

Other Enhancements

• Increasing Initial Window (IIW)

• Fast Start

• Teaching TCP to Ignore Transmission Errors

• Split TCP

• TCP Peach

Split TCP

TCP Peach

• Sudden Start

• Congestion Avoidance

• Fast Retransmit

• Rapid Recovery

Sudden Start

Rapid Recovery (Link Error)

Rapid Recovery (Congestion)