TCP Westwood: Efficient Transport for High-speed

wired/wireless Networks

2009

TCP Westwood (Mobicom 2001)

Key Idea:• Enhance congestion control via the Rate

Estimate (RE) Samples are determined from ACK inter-arrival times and info in ACKs regarding amounts of bytes deliveredEstimate is computed at the sender by sampling and exponential filtering

• RE is used by sender to properly set cwnd and ssthresh after packet loss (indicated by 3 DUPACKs, or Timeout)

Rate Estimation (BE-> RE)

• Ideally, would like to determine the connection fair share of the bottleneck bandwidth

• Since fair share is difficult (to define or determine), we instead estimate the achieved rate: Rate Estimate (RE)

Receiver

Sender

Internet

Bottleneck

packets

ACKs

measure

• First TCPW version used a “bandwidth like” estimator (BE) given by:

“Original” Rate estimation (BE-> RE)

)/( 1 kkkk ttdb

tk-1 tk

dk

2

1 11

kkkkkk

bbBEBE

sample

exponential filter

k

kk t

t

2

2filter gain

RE/BE Estimation is similar to Keshav Packet Pair estimation

TCP Westwood: the control algorithm

• TCPW Algorithm Outline:When three duplicate ACKs are detected:

set ssthresh=BE*RTTmin (instead of ssthresh=cwin/2 as in Reno)if (cwin > ssthresh) set cwin=ssthresh

When a TIMEOUT expires:

set ssthresh=BE*RTTmin (instead of ssthresh=cwnd/2 as in Reno) and cwin=1

Note: RTTmin = min round trip delay experienced by the connection

7

Extensions to TCP

• End-to-end congestion control

Selective acknowledgements: TCP SACK

Delay-based congestion avoidance: TCP Vegas

TCP-Friendly Rate Control: TFRC

Binomial Congestion Control

• Router - assisted Congestion Control

Active Queue Mgmt. (AQM) and RED

Explicit congestion notification: ECN

eXplicit Control Protocol: XCP

• Discriminating between congestion losses and other losses: cross-layer signaling and guesses losses

Related TCP + Bdw estimation work• Bandwidth estimate used by Vegas and Keshav PP

• TCP Vegas monitors Bdw and RTT to infer the bottleneck queue; then, from queue it derives feedback to congestion window

Expected = window size / round trip time

Actual = acks / round trip time

Queues (bottleneck) ~ expected - actual

• Keshav’s Packet Pair scheme also monitors bandwidth to estimate the bottleneck queue and compare to common queue target; it adjusts source rate

• Main difference: TCP Westwood does not need common queue target; it enforces fair share instead.

TCP Westwood Benefits

What do we gain by using RE “feedback” in addition to packet loss?

(a) ability to distinguish random loss from buffer loss, i.e. better performance with random loss (ie, loss caused by random errors as opposed to overflow)

-- exponential filtering

(b) using RE to estimate bottleneck bdw during slow start

-- estimate the achieved rate

TCPW and random loss

• Reno overreacts to random loss (cwin cut by half)• TCPW less sensitive to random loss

(1) a small fraction of “randomly” lost packets minimally impacts the rate estimate RE

(2) Thus, cwin = RE x RTT remains unchanged

• As a result, TCPW throughput is higher than Reno and SACK in presence of random loss

TCPW in presence of random loss: Analysis and Simulation

TCPW in a wireless lossy environment• Efficiency: Improvement significant on high (Bdw

x Length) paths

• Fairness: better fairness than RENO under varying RTT

• Friendliness: TCPW is friendly to TCP Reno

D a t a G e n e r a l

D a t a G e n e r a l

Base StationW ireless

Host

In te rnet In ternet

TC P

long p ropaga tion tim e

link e rro rs

NASA Workshop Demo(From Steve Schultz, NASA)

Internet Throughput Measurement

TCPW Friendliness

• Friendliness: fairness across different TCP flavors“Friendly share” principle: TCPW is allowed to recover the bandwidth wasted by NewReno because of “blind” window reduction

• TCPW original RE (BE) filter has Friendliness Problem….

5 connections total (TCPW + RENO) ; Average throughput per connection is shown in the next slide

Lossy link (1% pkt loss) TCPWReno

TCPW & Reno friendliness

Average TCPW & Reno throughputs vs. TCPW/Reno mix (5 connections total)

No link errors TCPWReno

0

0.2

0.4

0.6

0.8

1

1.2

1.4

Ave

rag

e T

hro

ug

hp

ut(

Mb

ps)

0 1 2 3 4 5No. of Reno connections

Ave

rag

e T

hro

ug

hp

ut(

Mb

ps

)

0

0.2

0.4

0.6

0.8

1

1.2

1.4

0 1 2 3 4 5No. of Reno connections

• Recall that the first TCPW version used a “bandwidth like” estimator (BE) given by:

TCPW original estimation (BE)

)/( 1 kkkk ttdb

tk-1 tk

dk

2

1 11

kkkkkk

bbBEBE

sample

exponential filter

k

kk t

t

2

2filter gain

TCPW Rate Estimation (TCP RE)

• Rate estimate (RE) is obtained by aggregating the data ACKed during the interval T (typically = RTT):

2

1 11

kkkkkk

bbRERE

sample

exponential filter

k

kk t

t

2

2filter gain

dk

T

dk-1

T

d

k

jTktjtb

tk T is the sample interval

BE overestimates fair rate ( = 2.5 Mbps)

TCPW BE Not friendly to NewReno!

TCPW RE or BE interaction with RENO

No errors (bottleneck gets saturated)

0

0.5

1

1.5

2

2.5

3

3.5

4

0 2 4 6 8 10 12 14

Bandw

idth

Est

imatio

n (

Mbps)

Time (sec)

BE RE

fair share

Errors (0.5%), no congestion

RE underestimates fair rate (=3.6 Mbps)

TCPW RE does not improve thruput!

0

0.5

1

1.5

2

2.5

3

3.5

4

0 2 4 6 8 10 12 14

Bandw

idth

Est

imatio

n(M

bps)

Time (sec)

BE RE

fair share*

(*) TCPW fair share > 50% because NewReno is incapable of getting 50%

One TCPW RE or BE and one Reno share a 5Mbps bottleneck

TCPW with adaptive filter (AF)• Neither RE or BE estimator are optimal for all

situationsBE is more effective in random loss RE is more appropriate in congestion loss (ie, buffer overflow)

• KEY IDEA: dynamically select the aggressive estimate (BE) or the conservative estimate (RE) depending on current channel status (congestion or random loss?)

• NEEDED: a “congestion measure” that gives us an idea of the most probable cause of packet loss (congestion or random)

• The Adaptive Filter actually provides a smooth transition from aggressive to conservative measure

TCPW AF: Sampling

• Adapting the size of sampling intervals to congestion level measure

Tk

Congestion: Tk growsTk

No Congestion: Tk = inter ACK

k

kTktjtj

T

d

ks

continuous adaptation

Rate sample

TCPW AF: Sampling (cont)

• The sample size Tk is continuously adjusted according to current congestion:

minmin

/)ˆ(*RTT

cwinhT

RTT

cwinRTTT kk

Max throughput assuming there is no

congestion in the networkactual achieved throughput

Sampleinterval Tk

Upon ACK Receipt

Severe Congestion: Tk --> RTT

Link Under Utilized: Tk --> 0 (ie, inter ACK intrv)

ABE Computes Congestion Level

TCP Westwood with Agile Probing: Handling Dynamic Large Leaky Pipes

-- Problem address Infocom 2004

• Leaky Pipes: packet loss due to error

Unjustified cwnd cut and premature Slow Start exit

• Large Pipes: Large capacity and long delay

Control scheme may not scale• Dynamic Pipes: Dynamic load/changing link

bandwidth (Due to change of technologies, e.g., 802.11, Bluetooth, GPRS)

Linear increase limits efficiency

Persistent Non-Congestion Detection (PNCD)

Persistent Non-congestion detected, Agile Probing invoked

Dominant flows leaveAt around 50 sec

Agile Probing

• Objective: Guided by ERE (Eligible Rate Estimate), converge faster to more appropriate ssthresh

adaptively and repeatedly resets ssthresh to ERE*RTTmin

Exponentially increase cwnd if ssthresh >cwnd

Linearly increase cwnd if ERE < ssthresh

Exit Agile Probing when packet loss is detected

Agile Probing

Lab Measurements Results (FreeBSD Implementation)

Persistent Non-Congestion Detection(PNCD) invokes Agile Probing after dominant flow left

Startup invokes Agile Probing

Summary

• Introduced the concept of Rate Estimation and related work

• Reviewed end-to-end estimation based congestion control methods

• Presented TCP Westwood, and the evolution of “fair rate” estimate to improve the performance; showed simulation results to evaluate the method

• Compared TCPW with other methods