34
Fast TCP Matt Weaver CS622 Fall 2007

Fast TCP

  • Upload
    airell

  • View
    33

  • Download
    0

Embed Size (px)

DESCRIPTION

Fast TCP. Matt Weaver CS622 Fall 2007. FAST TCP: Motivation, Architecture, Algorithms, Performance. David X. Wei, Student Member, IEEE, Cheng Jin, Steven H. Low, Senior Member, IEEE, and Sanjay Hegde. Abstract. - PowerPoint PPT Presentation

Citation preview

Page 1: Fast TCP

Fast TCP

Matt WeaverCS622 Fall 2007

Page 2: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

FAST TCP: Motivation, Architecture, Algorithms, Performance

David X. Wei, Student Member, IEEE, Cheng Jin, Steven H. Low, Senior Member, IEEE, and Sanjay Hegde

Page 3: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Abstract FAST TCP is a congestion control algorithm that attempts to solve

the problems of congestion control. This paper covers:

The algorithm itself. Performance metrics

3

Page 4: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Background “Congestion control is a distributed algorithm to

share network resources among competing users.”

A difficult problem to solve...Resource needs vary, depending on time of day.Available resources is usually static.

4

Page 5: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

FAST TCP FAST is a recursive acronym:

FAST AQM Scalable TCP○ AQM: Active Queue Management○ TCP: Transmission Control Protocol (duh)

5

Page 6: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Current Issues As congestion is monitored, current algorithms

slow down monitoring as packets are dropped, the average sending rate depends on low loss probability.High data transmission rates are required for low

loss.Usually lower than WiFi can support.

6

Page 7: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Solution FAST TCP uses queues to store a constant

number of packets.If too few packets are queued, the sending rate

increases.If too few, the rate decreases.

7

Page 8: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Congestion ControlCurrent TCP congestion control algorithm (aka Reno). At the packet level, linear increase by one packet per roundtrip time

(RTT) is too slow, and multiplicative decrease per loss event is too drastic.

At the flow level, maintaining large average congestion windows requires an extremely small equilibrium loss probability.

At the packet level, oscillation in congestion window is unavoidable because TCP uses a binary congestion signal (packet loss).

At the flow level, the dynamics is unstable, leading to severe oscillations that can only be reduced by the accurate estimation of packet loss probability and a stable design of the flow dynamics.

8

Page 9: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Motivations Two levels of design:

The flow level (macroscopic) covers:○ QoS○ Stability○ etc

Packet level (microscopic) covers:○ The same goals, but focused on end to end.

Reno suffered because higher level control was considered after the micro level.

9

Page 10: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Calculations Congestion and

utility: U calculates utility

for each stakeholder (user) at a given flow.

10

Page 11: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Calculations Equilibrium (FAST):

11

Equilibrium (Reno):

Page 12: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Calculations “A key departure of our model from those in the

literature is that we assume that a source’s send rate, defined as xi(t) :=wi(t)=Ti(t), cannot exceed the throughput it receives. “

12

Page 13: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Dynamic Structure The weakness of current schemes versus FAST is

shown for large window sizes.

13

Page 14: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Equilibrium Equilibrium measures congestion consistency.

14

Page 15: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Differences Though all of the aforementioned algorithms

look different at the packet level, they actually have similar structures at the flow and equilibrium levels.

15

Page 16: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Performance Test 1 To test performance, packet data is pushed

through a semi-articial network.Identical sender and receiver boxes, running

dummynet on FreeBSD.Emulated router.Dummynet running:

○ Paths with RTTs of 50, 100, 150, and 200ms.○ Second path with a bottleneck capacity of 8M/s and a

buffer size of 2,000 packets shared by all the delay pipes.

16

Page 17: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Results Dynamic state I:

Small flows, large windows Dynamic state II:

Larger flows

17

Page 18: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Dynamic State I

18

Page 19: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

19

FAST vs Reno I

Throughputkbps

Queue(avg)# of pkts

Sec

Page 20: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Dynamic State II

20

Page 21: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Dynamic State II

21

Page 22: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

22

FAST vs Reno II

Page 23: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

23

BIC

Page 24: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Performance Test 2 Dummynet tests are limited to a single

bottleneck and the same protocols. NS-2 Simulation run in lab:

Same algorithm.Noise added to eliminate phase artifacts.

24

Page 25: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

25

Page 26: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

26

Page 27: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

27

Page 28: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

28

Page 29: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

29

Page 30: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

30

Page 31: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

Conclusion Because of the use of queues, FAST TCP can

handle lower transmission rates. The paper also covers some simulated

scenarios (too lengthy to cover properly here).

31

Page 32: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

32

Caveat Emptor Possibly biased research:

Jin Cheng, Steve Low, and David Wei (the authors) patented and market the FAST TCP algorithm.

FAST TCP implementation sold as FastSoft Aria (a 1 U rack mountable hardware solution). http://www.fastsoft.com/

Ao Tang proposed that these measurements were somewhat misleading in another paper.

Page 33: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

33

Other CC Algorithms BIC TCP Compound TCP CUBIC H-TCP High Speed TCP HSTCP-LP Hybla New Reno Tahoe TCP-Illinois TCP-LP TCP-SACK TCP-Veno Westwood Westwood+ XCP YeAH-TCP

Page 34: Fast TCP

22/04/23 FAST TCP: Motivation, Architecture, Algorithms, Performance

34

Related Work Tang, A., Wang, J., Low, S. H., and Chiang, M. 2007. Equilibrium of heterogeneous congestion

control: existence and uniqueness. IEEE/ACM Trans. Netw. 15, 4 (Aug. 2007), 824-837. DOI= http://dx.doi.org/10.1109/TNET.2007.893885

Ma, J., Ruutu, J., and Wu, J. 2000. An enhanced TCP mechanism—fast-TCP in IP networks with wireless links. Wirel. Netw. 6, 5 (Nov. 2000), 375-379. DOI= http://dx.doi.org/10.1023/A:1019118421144

Gu, Y., Hong, X., and Grossman, R. L. 2004. Experiences in Design and Implementation of a High Performance Transport Protocol. In Proceedings of the 2004 ACM/IEEE Conference on Supercomputing (November 06 - 12, 2004). Conference on High Performance Networking and Computing. IEEE Computer Society, Washington, DC, 22. DOI= http://dx.doi.org/10.1109/SC.2004.24

Grieco, L. A. and Mascolo, S. 2004. Performance evaluation and comparison of Westwood+, New Reno, and Vegas TCP congestion control. SIGCOMM Comput. Commun. Rev. 34, 2 (Apr. 2004), 25-38. DOI= http://doi.acm.org/10.1145/997150.997155