Satellite TCP Kenny - Simon Fraser Universityljilja/ENSC835/Fall03/News/Presentations/Satellite_TC… · Satellite VSAT VSAT VSAT VSAT µç»° ... µç»° ´«Õæ ½»»»»ú

ENSC 835 project (2002)TCP performance over satellitelinks

Kenny, Qing [email protected], Hui Zhang

[email protected]

Sep, 23 2003 2

Road map

n Introduction to satellite communicationsn Simulation implementation

n Window sizen Maximum segment sizen Initial windows sizen Selective acknowledgement algorithm§ TCP burst

n Conclusionn References

Sep, 23 2003 3

Introduction to satellitecommunications

• Satellite communication is a type of the wirelesscommunications technologies. It utilizes satellites toretransmit the wireless signal, and to connect withthe multiple earth station.

Sep, 23 2003 4

Introduction to satellitecommunications

Geosynchronous Equatorial Orbit

(From geo = earth + synchronous = moving at the same rate.)

Sep, 23 2003 5

On board GEO satellite

Bent pipe GEO satellite On board switch GEO satellite

HUB

Fax

Telephone

Public Switch

Satellite

VSAT

VSAT

VSAT

VSAT

µç»°

µç»°

´«Õæ

½»»»»ú

Star network (switch at hub) Mesh network (switch at satellite)

Sep, 23 2003 6

Why do we choose this project?

n Commercial satellite companies (e.g., Loral, Hughes,Lockheed Martin) have announced plans to buildlarge satellite systems to provide broadband dataservice.

n Our simulation may help improve TCP performanceover long delay and error prone channels.

Sep, 23 2003 7

Project objective

n Implement an combination of four approaches toenhance the performance over satellite links couplingwith the long delay and high Bit Error Ratecharacteristics.n Effect of window size RFC1323n Effect of Initial windows size RFC2581n Effect of Maximum segment size RFC2488n Comparison of different TCP algorithms RFC 2018

n Extend the authors’ knowledge of TCP burst problemrelated to on board switch in GEO satellites.n TCP burst

Sep, 23 2003 8

Window size

n Large delay*bandwidth productW=B*RTT

This product defines the amount of data aprotocol should have “ in flight”.

Sep, 23 2003 9

Window size

n The original TCP standard limits the advised windowsize by only assigning 16bits of header space for itsvalue. (RFC793). Hence the advised window size canbe no more than 64Kbytes.

time trip roundCWND

=throughput

kbpsbytesms

kbyte900sec/000,112

58564

≈≈=

Sep, 23 2003 10

Window size

n Simulation scenario

n Parameters:n choose various window size: 16, 32, 64, and 128.n T1 link: 1.544Mbps.n Set the receiver and sender buffer size greater

than the delay bandwidth product, so that we canexamine how window size affect on TCPthroughput: 120.

n Application: FTP

Sep, 23 2003 11

Window size (Result 1)

• Brief Analysis:Larger windowsize can helpimprove thethroughput.

Sep, 23 2003 12

Window size (Result 2)

• Brief AnalysisØ Multiple long lived

connections withsmall window sizecan still fully utilizethe channel.ØThree 40kbytes

window connectionscan almost fullyutilize the T1channel.

Sep, 23 2003 13

Window size -Discussion

n Advantages:n It is ideal for the connections that transmit

big files such as FTP applicationn Disadvantages:

n large window size can lead to more rapiduse of the TCP sequence space.

n large window size will also increase themultiple packets loss possibility.

Sep, 23 2003 14

What about short lived connection?

n Slow start is a safe guard against transmitinappropriate amount of data into thenetwork.

n However, Slow start is particularly inefficientfor short lived connection (Telnet) in largebandwidth-delay product network.

Sep, 23 2003 15

Initial congestion window size

n By increasing the initial CWND, more packetsare sent during the first RTT.

n Trigger more ACKs, allowing congestionwindow to open more rapidly.

Sep, 23 2003 16

Maximum Segment Size

Maximum segment size (MSS):

MSS = MTU – TCP header – IP header

n MTU (or the maximum IP packet size):Maximum Transmission Unit

Sep, 23 2003 17

Initial CWND and MSS

q Simulation scenario

n Parameters:n Set the initial window size: 1, 2, and 4.n Set packet size: 576, 1152, and 1728.n Set the advised window size: 128.n Application : telnetn Other parameters are default.

Sep, 23 2003 18

• Brief Analysis:Using larger initial window size can help reduce the slow start period,so allows sender send more packets at the same period of time.

Initial window size (Result 1)

Sep, 23 2003 19

Maximum Segment Size (Result)

• Brief Analysis:Choosing suitablemaximum segmentsize can improveTCP throughput.

Sep, 23 2003 20

Initial CWND and MSS-Discussion

n Advantages:n It is ideal for the short lived connections

such as Telnet application.n Disadvantages:

n Make traffic burstn Increase unnecessary drops for bigger

packet size

Sep, 23 2003 21

Comparison of different TCPalgorithms

n New Challenge for TCP performance:n Large window size will also increase the

multiple packets loss possibility.n Large MSS will easily be corrupted in

wireless link.

n The need to evaluate the error correct abilityof different TCP flavors.

Sep, 23 2003 22

Comparison of different TCPalgorithms

n Simulation scenario

n Parameters:n Add lose model, use different bit error rate : 10e-7, 10e-6,

10e-5, 10e-4, and 10e-3.n Compare TCP Sack with other TCP algorithms such as Reno

and Vegas.

Sep, 23 2003 23

Result

• Brief Analysis:TCP Sack performsbetter than Reno.However, when biterror rate is from5*10e-7 ~ 10e-3,Vegas is better thanboth Sack and Reno.

0

200

400

600

800

1000

1200

1400

Bit erro rate

Thr

ough

put

(kbp

s)

SackRenoVegas

10e-7 10e-6 10e-5 10e-4 10e-3

Sep, 23 2003 24

On board GEO satellite

Bent pipe GEO satellite On board switch GEO satellite

HUB

Fax

Telephone

Public Switch

Satellite

VSAT

VSAT

VSAT

VSAT

µç»°

µç»°

´«Õæ

½»»»»ú

Star network (switch at hub) Mesh network (switch at satellite)

Sep, 23 2003 25

TCP Burst

n Larger window size and initial windows sizewill results in large burst.

n On board switch has limited buffer.

n New Challenge for TCP performance:n TCP burst Vs limited buffer sizen Suitable buffer size

Sep, 23 2003 26

TCP Burst

n Simulation scenario:

n Parameters:n Set various buffer size: 30, 34, 40, and 50.n Set the window size: 128.n T1 link : 1.544Mbps.n Other parameters are default.

1.5 44Mbps,3.1Mbps 3.1Mbps

source sinkrouter1 router2

Sep, 23 2003 27

TCP Burst

• Brief Analysis:§ Increasing the buffer size will not alleviates the TCP burst§ 2. After the slow start period, the queue occupancy decrease.

Sep, 23 2003 28

TCP Burst

• Brief Analysis:Increasing the buffer size will not improve the TCP performance.If the buffer size is approximately greater than 1/2 of the windowsize when using basic ACK (accordingly, 1/3 when using delayedACK), the throughput will reduce rapidly.

with basic acknowledgment with delayed acknowledgment

Sep, 23 2003 29

Conclusion

Using a larger initial window can improve thethroughput, especially for short transfers

Initial window

TCP Sack is better than Reno under the error pronechannel. However, Vegas performs even betterthan Sack if the bit error rate is large

Comparisonof three TCPalgorithms

Larger maximum segment size will improve thethroughput, however, it may cause link congestionor router overload

Maximumsegment size

Burst has a severe influence on TCP performanceTCP Burst

Larger window size can improve the performanceWindow size

OutcomeExperiment

Sep, 23 2003 30

Conclusion

n For FTP application in satcom, implementSACK or Vegas with larger window size cangreatly improve the TCP throughput.

n For Telnet application in satcom, it is ideal toimplement SACK with larger initial windowand MSS.

n The ideal buffer size of on board satelliterouter is ½ of the window size (standardACK), 1/3 of the window size (delay ACK).

Sep, 23 2003 31

Reference

n [1] Van Jacobson. Congestion Avoidance and Control. In ACM SIGCOMM, 1988.n [2] J. Mo, R. J. La, V. Anantharam, and J. Warland, ``Analysis and comparison of TCP Reno

and Vegas,'' Proceedings of the Conference on Computer Communications (IEEE Infocom),New York, Mar. 1999.

n [3] L. Brakmo, S. O'Malley, and L. Peterson. TCP Vegas: New techniques for congestiondetection and avoidance. In Proceedings of the SIGCOMM '94 Symposium (Aug. 1994)pages 24-35

n [4] K. Fall and S. Floyd, "Simulation-based comparisons of Tahoe, Reno, and SACK TCP,"SIGCOMM Computer Communication Review, 26(3), July 1996

n [5] A Simulation Study of Paced TCP, Joanna Kulik, Robert Coulter, Dennis Rockwell, andCraig Partridge, September 1999

n [6] [RFC 2488] Enhancing TCP over Satellite Channels, Mailman, and D.Glover January1999.

n [7] Larry L.Peterson and Bruce S.Davie. Computer networks: A system approach. MorganKaufman, 1996.

n [8] [RFC 2581] TCP Congestion Control, M. Allmanm, V. Paxson, W. Stevens, April 1999.n [9] [RFC 1122] Transport Layer TCPn [10] [RFC 1323] TCP Extensions for High Performance, V.Jacobson, R.Braden, D.Borman,

1992