18
Modeling of Web/TCP Transfer Latency Yujian Peter Li January 22, 2004 M. Sc. Committee: Dr. Carey Williamson Dr. Wayne Eberly Dr. Elena Braverman Department of Computer Science, University of Calgary

Modeling of Web/TCP Transfer Latency

Embed Size (px)

DESCRIPTION

Modeling of Web/TCP Transfer Latency. Yujian Peter Li January 22, 2004 M. Sc. Committee: Dr. Carey Williamson Dr. Wayne Eberly Dr. Elena Braverman. Department of Computer Science, University of Calgary. Outline. Motivation and Objectives. TCP Overview and Related Work. - PowerPoint PPT Presentation

Citation preview

Page 1: Modeling of Web/TCP Transfer Latency

Modeling of Web/TCP Transfer LatencyYujian Peter Li January 22, 2004

M. Sc. Committee: Dr. Carey WilliamsonDr. Wayne EberlyDr. Elena Braverman

Department of Computer Science, University of Calgary

Page 2: Modeling of Web/TCP Transfer Latency

2

Outline

TCP Overview and Related Work

The Proposed TCP Transfer Latency Model

Model Validation by Simulation

Extending the Proposed Model to CATNIP TCP

Conclusions

Motivation and Objectives

Page 3: Modeling of Web/TCP Transfer Latency

3

Motivation

Web response time Highly dominated by TCP performance

Understanding the sensitivity of TCP to network conditions helps to improve TCP performance

No work on modeling CATNIP TCP

Page 4: Modeling of Web/TCP Transfer Latency

4

Objectives

To survey and compare existing TCP models

To develop an accurate model for short-lived TCP flows

To model CATNIP TCP

Page 5: Modeling of Web/TCP Transfer Latency

5

SYNSYN/ACK

ACK

FIN FIN/ACKACK

TCP OverviewCharacteristics

Reliable, in-order byte stream

Flow control Connection-oriented Congestion Control

Web Browser Web Server DATA

Page 6: Modeling of Web/TCP Transfer Latency

6

TCP OverviewCongestion Control

• When intermediate nodes (routers) become overloaded, the condition is called congestion.

• The mechanisms to solve the problem are called congestion control.

Page 7: Modeling of Web/TCP Transfer Latency

7

TCP Overview – Congestion ControlSlow Start & Congestion Avoidance

Slow start: cwnd=cwnd+1 for every received ACK

Congestion avoidance: cwnd = cwnd + 1/cwnd

Page 8: Modeling of Web/TCP Transfer Latency

8

Related Work

TCP Steady State Throughput Model [Padhye et al. 1998]

TCP Response Time Models Cardwell-00 Model [Cardwell et al. 2000] Padhye Model [Cardwell et al. 1998] Cardwell-98 Model [Cardwell et al. 1998] Sikdar Model [Sikdar et al. 2001]

Page 9: Modeling of Web/TCP Transfer Latency

9

The Proposed TCP Response Time ModelAssumptions

Bernoulli packet loss model, i.e., packet is independently lost with fixed probability p Congestion avoidance algorithm ignored, i.e., cwnd always increases by one upon receiving one ACK (exponentially)

Packet loss can be via RTO or triple duplicate ACKs

The effect of delayed ACK, Tdelay, is added when necessary

Page 10: Modeling of Web/TCP Transfer Latency

10

The Proposed Model (Cont’d)Congestion Window Evolution

iloss

iss

lastCycle

iTTET

][

1

Page 11: Modeling of Web/TCP Transfer Latency

11

Simulation ExperimentsNetwork Topology

Page 12: Modeling of Web/TCP Transfer Latency

12

Factor Levels

Transfer Size 1KB, 4KB, 8KB, 16KB, 32KB, 50KB, 64KB, 90KB, 110KB, 128KB, 160KB, 180KB, 200KB

Packet Loss Probability 1%, 3%, 5%, 8%, 10%

Simulation ExperimentsMetric & Experimental Factors

Performance Metric: Data Transfer Time, the time from when the sender sends the first packet until the time when the sender receives the ACK of the last data packet.

Experimental factors and levels

Page 13: Modeling of Web/TCP Transfer Latency

13

Simulation ResultsShort-lived Flows

 ( p=3%) (p=10%)

0

1

2

3

4

5

6

7

8

0 5 10 15 20 25 30 35 40

Data Transfer Size (Packets)

Tran

sfer

Tim

e (s

ec)

Simulated

Proposed

Sikdar

Cardwell-000

1

2

3

4

5

6

7

8

0 5 10 15 20 25 30 35 40

Data Transfer Size (Packets)

Tran

sfer

Tim

e (s

ec)

Simulated

Proposed

Sikdar

Cardwell-00

Page 14: Modeling of Web/TCP Transfer Latency

14

CATNIP TCP

  C. Williamson and Q. Wu : “A Case for Context-Aware TCP/IP”.

ACM Performance Evaluation Review, Vol. 29, No. 4, pp. 11-23, March 2002. Convey application-layer context information to TCP/IP Not all packet losses created equal

IP

TCP

HTTPDocument Size

Packet Loss Priority

Page 15: Modeling of Web/TCP Transfer Latency

15

CATNIP TCP v.s. Partial CATNIP TCP

Packets First three

Last three

cwnd<3

CATNIP

Partial CATNIP

Page 16: Modeling of Web/TCP Transfer Latency

16

0

10

20

30

40

50

60

70

1 2 3 4 5 6 7 8 9 10

Transfer Time (sec)

Fre

qu

ency

(%

)

Partial CATNIP

CATNIP

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 8

Transfer Time (sec)

Cu

mu

lati

ve F

ract

ion

Partial CATNIP

CATNIP

0

10

20

30

40

50

60

70

1 2 3 4 5 6 7 8 9 10

Transfer Time (sec)

Fre

qu

ency

(%

)

Partial CATNIP

CATNIP

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 8

Transfer Time (sec)

Cu

mu

lati

ve F

ract

ion

Partial CATNIP

CATNIP

0

10

20

30

40

50

60

70

1 2 3 4 5 6 7 8 9 10

Transfer Time (sec)

Fre

qu

ency

(%

)

Partial CATNIP

CATNIP

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 8

Transfer Time (sec)

Cu

mu

lati

ve F

ract

ion

Partial CATNIP

CATNIP

CATNIP TCP v.s. Partial CATNIP TCP

p=3%

p=5%

p=10%

PDF CDF

Page 17: Modeling of Web/TCP Transfer Latency

17

Modeling Partial CATNIP TCP Short-lived Flows

 ( p=3% p’=0%) (p=10% p’=0%)

0

1

2

3

4

5

6

7

0 5 10 15 20 25 30 35 40

Data Transfer Size (Packets)

Tran

sfer

Tim

e (s

ec)

TCP Reno (Simulated)

Partial CATNIP (Simulated)

Partial CATNIP (Analytical)

0

1

2

3

4

5

6

7

0 5 10 15 20 25 30 35 40

Data Transfer Size (Packets)

Tran

sfer

Tim

e (s

ec)

TCP Reno (Simulated)

Partial CATNIP (Simulated)

Partial CATNIP (Analytical)

Page 18: Modeling of Web/TCP Transfer Latency

18

Conclusions

The proposed TCP latency model fits the simulation results better than earlier models.

The differences between Partial CATNIP and CATNIP are minimal when p<10%.

Partial CATNIP TCP model matches the simulation as well.

Partial CATNIP TCP improves TCP latency compared to TCP Reno. For short-lived flows, Partial CATNIP TCP is about 10% faster than TCP Reno in most cases.

CATNIP TCP is a suitable approach to improve TCP Performance.