Postech DP&NM Lab Freeze-TCP: a true end-to-end TCP enhancement mechanism for mobile environments Goff, T.; Moronski, J.; Phatak, D.S.; Gupta, V. INFOCOM

Postech Postech DP&NM LabDP&NM Lab

Freeze-TCP: a true end-to-end TCP enhancement mechanism for mobile envi

ronments

Goff, T.; Moronski, J.; Phatak, D.S.; Gupta, V. INFOCOM 2000

Lee Hyo Jin2001 Fall Mobile Networks 발표자료

Nov/28/2001

Prof. Young-Joo Suh

Freeze-TCP (2)Postech Postech DP&NM LabDP&NM Lab

Reference• Tom Goff et el, "Freeze-TCP: A True End-to-End TCP Enh

ancement Mechanism for Mobile Environments," INFOCOM'00.

• K. Brown and S. Singh, “M-TCP: TCP for Mobile Cellular Networks,” ACM Computer Communications Review (CCR), vol. 27, no. 5, 1997.

• Ajay Bakre and B.R. Badrinath, “I-TCP: Indirect TCP for mobile hosts,” Tech. Rep., Rutgers University, May 1995,


Contents

• Introduction.• Requirement• Key concepts.• TCP window management.• Introduce to existing solutions.• Details of Freeze-TCP.• Experimental result.• Conclusion and Discussion.


Introduction

• Need to optimize TCP for mobility.• Not true end-to-end scheme.

– Intermediaries. ( like Base Stations )• To monitor the TCP traffic and participate in flow control to

enhance TCP performance.

– Not applicable when IP payload is encrypted.(IPSEC)• Security associations between sender and receiver.

• Require changes TCP/IP code at intermediate node.– It is difficult for mobile clients to inter-operate with the

existing infrastructure.


Requirements

• True end to end scheme.• Interoperate existing infrastructure.

– TCP code must change in mobile client (MH)

• Need to performance enhancement.

We need a new scheme.


Key Concepts

• No help with base stations in hand-off.• To detect an impending handoff at client.( MH )• ZWA(MH): zero window advertisement.

• ZWP (FH) : zero window probes.

• TR-ACKs : Triplicate acks.

• True end to end semantics.

• Performance enhancement.


TCP window management -1

• The window size – minimum of receiver’s advertised buffer size

– perceived network congestions.

• The receiver run out of its buffer-space and advertise a window size of zero. ( ZWA )

• The sender should freeze all retransmit-timers and enter a persist-mode on seeing ZWA.



Ack1 win4

12

43 4 5 6

8

DATA3 ~ 6 win4

Data1 win4

Ack6 win0Ack6 win4

9

10 11 12 13

DATA10 ~13 win4 ZW

A

sender

receiver



• ZWP

• Sending probes until the receiver’s window opens up.

• Sender want to knows receiver’s window opened or not, in advance.

• Interval – exponential back-off until it reaches 1 minute

– remains constant after 1 minute.

• Receiver responds to a ZWP with a non-zero window size.

• Sender will continue transmission using a window size consistent with the advertised value.



Ack1 win4

12

43 4 5 6

8

DATA3 ~ 6 win4

Data1 win4

Ack6 win0Ack6 win4

9

10 11 12 13

DATA10 ~13 win4 ZW

A

ZWP



8Ack6 win0

9

10 11 12 13 DATA10 ~13

win4

ZWP

Probe response (win4)

Original ack


Existing Solutions

• SNOOP • I-TCP ( Indirect TCP )• EBSN ( Explicit bad state notifications )• Delayed dupacks• M-TCP


I-TCP• Split the connection

– FH-BS : Standard TCP.– BS-MH : Standard TCP ,Optimi

zing protocol.(MTCP)

• Retain a little RTT – Fast recovery about cwnd size d

egradations.

• Need to exchange the status information – Long delay time.– MSR buffer size is small. (to re

duce handoff time)– MSR : Mobility Support Router

s.

MH

FH

MH

MH socket(mhaddr, mhport, msr1addr, msr1port)

MSR1or 2 mhsocket(msr1addr, msr1port, mhaddr, mhport)

MSR 1 MSR 2

MSR1or 2 fhsocket(mhaddr, mhport, fhaddr, fhport)

FH socket(fhaddr, fhport, mhaddr, mhport)

Regular TCP

Wireless TCP


EBSN

• Explicit bad-state notifications.• BS sends an EBSN to sender when each failed

attempt to send a packet to a MH.• On receipt of each EBSN, the sender reset

retransmission timer to original value.• Prevent the sender from dropping congestion

window.


M-TCP (1)

• Performance enhancement during hand-off.

• Low BER and Frequent disconnections.

• 3 level hierarchy structure.– Reduce MSS functions

– No need to exchange the status info moving MSS in the same SH domain.

High-speed Network

SH SH

MH

MSS

Cell

SH : Supervisor Host

MSS : Mobile Support Station

MH : Mbile Host


M-TCP (2)

• End to end TCP semantics.– TCP connection is split at the BS

– The SH does not send an ack FH unless BS has received an ack from MH.

FH

MH

BS


M-TCP (3)

• TCP Persist Mode– When the positive window advertisement is received, s

ender exits persist mode.

– Retain RTO and congestion window size.

• Need help from BS.– BS detect disconnection or packet loss.

– BS withholds ack for last one byte.

– Re-packetization penalty at sender.

– This ack uses to send to zero window advertisement at hand off.


TPC Performance

enhancement degradation

M-TCP Good Re-packetization overhead at sender

SNOOP

I-TCP( MTCP )

Handle

Bit-error

Frequent hand-off or disconnections

Delayed dupacks

Single packet

losses

Actual congestion losses

EBSN Significant duration or burst error

Random

error


Picture of Freeze-TCP

Fixed Host

(Sender)

ZWA

ZWP

BS

BS

MH

MH MH

MH

Connection

Probe res


ZWPFreeze-TCP (2)

• ZWP – ZWA force the sender into the ZWP (persist) mode.

– To prevent it from dropping its congestion window.

– To send ZWPs until the receiver’s opens up

– The interval grows exponentially (exponential back off ) until it reaches 1 minute.

– ZWP reponse does not have receiver’s advertisement window size.


Warning PeriodFreezeTCP (3)

• Warning period.– How much in advance of the disconnection should the

receiver start advertising ZWA?

– Ideally, long enough to ensure that exactly one ZWA get across to the sender.

– Longer : idle time prior to disconnections

– Small : sender’s congestion window to drop.

– RTT is reasonable. ( ref : Experimental result )

– Only useful if a disconnection occurs while data is being transferred.


TR-ACK -1Freeze-TCP (3)

• Triplicate Reconnection ACKs– ZWPs are exponentially backed off.

– The possibility of idle time after reconnections.

– To avoid this idle time, TR-ACKs implements.

– Effect of standard TCP.


TR-ACK

ZWP

ZWP

sender

receiver

Receiver window open

Sending again


Estimate performance -1Freeze-TCP (4)

• Idle period avoided.– W • ts ≥ RTT , W ≥ RTT / ts

: ts ≈ packet-size / band width , W : sender window size

Receiver

SenderRTT

W unACKed packets can be sent

ts


Estimate performance -2Freeze-TCP (5)

• Increased throughput.


Experimental result

• Modifying the Linux 2.1.101 TCP source code.• Emulate the mobile host in a PC.• Freeze-TCP is not worsen performance by a

noticeable amount.


Conclusion and Discussion -1

• To enhance TCP performance in the present of disconnections and reconnections.

• True end-to-end signaling scheme.• Unnecessary intermediaries’s help.• Easy changing TCP code at receiver side • Easy to implement. • Almost no overheads and tradeoffs.• Complete inter-operability with existing infrastruc

ture is guaranteed.


Conclusion and Discussion -2

• Full rate with old window size on entering new unknown environment or not.

• Needs at receiver to predict impending disconnections. ( pro-active action/simulations )

Documents

Postech DP&NM Lab Freeze-TCP: a true end-to-end TCP enhancement mechanism for mobile environments Goff, T.; Moronski, J.; Phatak, D.S.; Gupta, V. INFOCOM