TCP Westwood:

1
TCP Westwood Enhanced Congestion Control for
Large Leaky Pipes
Prepared for NASA Workshop on June 25,
2001 Mario Gerla, Giovanni Pau, M. Y. Sanadidi,
and Ren Wang UCLA Computer Science
Department Previous Contributors Saverio
Mascolo, Politecnico di Bari Claudio Casetti,
Politecnico di Torino Scott Lee, UCLA Computer
Science Department
2
Outline

• Concept
• Goal
• Algorithm
• Applications
• Performance
• Bandwidth Estimation Accuracy
• Throughput Gain
• Wireless experiments
• Analytic Model
• Internet Measurements
• TCPW Fairness/Friendliness
• Fairness among TCPW Connections
• Friendliness to TCP Reno Connections
• Future Research

3
TCP Congestion Control
4
Goal

• Enhance Congestion Control Using Bandwidth Share
  Estimate (BSE)
• BSE is computed at the sender by sampling and
  exponentially averaging an estimate of the
  instantaneous bandwidth share used by the
  connection
• Bandwidth samples are determined from ACK
  inter-arrival times and info in ACKs

5
Algorithm

• When three duplicate ACKs are detected
• set ssthreshBSERTT
• (instead of ssthreshcwin/2 as in Reno)
• if (cwin gt ssthresh) set cwinssthresh
• When a TIMEOUT expires
• set ssthreshBSERTT set cwin1
• (instead of ssthreshcwnd/2 as in Reno)

6
Applications
TCPW Is Particularly Effective Over Large Leaky
Pipes
7
Bandwidth Estimation Accuracy
Performance Evaluation (1)
BSE Accuracy (With Concurrent On/Off UDP Traffic)
8
Performance Evaluation (2)
Throughput Gain
9
Performance Evaluation (3)
Throughput Gain (Contd)
10
Performance Evaluation (4)
Wired/Wireless Model
lbg
2-State Markov Model for Burst Error
Characterization
Impact of Blackout Duration
11
Performance Evaluation (5)
LEO Scenario (Simulation)
200 msec RTT LEO Link Normal Operation Packet
Error Rate is 0.1 Handoff Blackout 100 msec 4
sec Inter-handoff period
Impact of Transmission Speed
12
Performance Evaluation (6)
Analytic Model

• W0(m) evolves as Markov chain

• The average throughput is

13
Performance Evaluation (7)
Throughput Gain (Analysis Validated By
Simulation)
45Mb/s link 70 msec RTT Router Buffer Size294
( Pipe Size)
14
Performance Evaluation (8)
Internet Measurements Testbed
15
Performance Evaluation (9)
Preliminary Internet Measurement Results
(One Direction Of The Connection To Brazil Was
Over Satellite Link )
16
Performance Evaluation (10)
NASA Workshop Demo (From Steve Schultz, NASA)
Internet Throughput Measurement
17
Fairness/Friendliness (1)
Why Is TCPW Fair?

• Two competing sessions
• Wi Ri (B/C RTT) for i 1,2
• In the increase phase windows grow at the same
  rate
• At overflow, windows scaled back proportionally
  to zero backlog line

18
Fairness/Friendliness (2)
Better Fairness When Pipe Lengths Vary
Fairness Among Long RTT And Short RTT Connections
Without and With RED
19
Fairness/Friendliness (3)
TCPW Fairness (Internet Measurements)
20
Fairness/Friendliness (4)
TCPW And TCP Reno Can Coexist
Throughput For A Connection Mix (Total of 5
Connections)
21
Future Research

• TCPW Stability Analysis
• Enhanced And Controlled Friendliness Tradeoffs
  Between Friendliness And Efficiency
• TCPW Internet And Internet II Measurements
• TCPW With NewReno Features
• BSE (Without TCP) For Adaptive Video Streaming