Modelling TCP Reno with Spurious Timeouts in Wireless Mobile Environments

1ICCCN 2003

Modelling TCP Reno with Spurious Timeouts in Wireless Mobile Environments

Shaojian Fu

School of Computer ScienceUniversity of Oklahoma.

Email: sfu@ou.edu

2ICCCN 2003

Introduction

Delay spikes are especially more frequent in today's wireless mobile networks than in traditional wired network, which will cause Spurious Timeouts (ST).

Previous studies on Modelling TCP performance over wireless networks focus on the impact of wireless random losses.

Spurious Timeouts must be considered explicitly to accurately

model the steady state sending rate and throughput of TCP.

Proposed an analytical model for the sending rate and throughput of TCP Reno as a function of packet error rate and characteristics of spurious timeouts.

3ICCCN 2003

Outline

The background on Delay Spike and Spurious Timeout provided.

Impact of this model on future transport protocol research.

Modelling assumptions.

Analytical model.

Validation of the model with simulation works.

Compare with previous model on TCP perfromance.

4ICCCN 2003

Examples of Delay Spikes

TimeNormal RTT

Round Trip Time

Delay spikes

5ICCCN 2003

Events causing delay spikes in wireless mobile environment

The handoff of a mobile host between cells requires the registration with the base stations.

The physical disconnection of the wireless link during a hard handoff.

Retransmission at Radio Link Control (RLC) layer, e.g. GPRS and CDMA2000.

Higher-priority traffic, such as circuit-switched voice, can preempt (block) the data traffic temporarily.

6ICCCN 2003

Spurious timeout illustrated

7ICCCN 2003

Sender’s congestion window

Spurious

Timeout Begin transmit new data

8ICCCN 2003

Previous models on TCP performance

Early TCP analytical models only consider slow start and congestion avoidance.

Recent models take into account the RTO timeout caused by random losses during transmission, such as the model proposed by Padhye et. al. (Referred as PFTK model in our paper).

Since Spurious Timeouts are not frequent in wired networks, they are considered to be a transient state, and thus cannot produce much impact on the steady state performance of TCP.

In wireless mobile environment, Spurious Timeouts are more frequent. They must be modelled explicitly to estimate the steady state performance of TCP more accurately.

9ICCCN 2003

A new analytical model considering the characteristics of Spurious Timeouts

Impacts of Spurious Timeouts are explicitly built into the analytical TCP performance model.

Stochastic analysis of the steady state sending rate and throughput of TCP Reno as a function of:

packet error rate, interval between long delays, duration of long delays.

The model proposed by Padhye et. al. (referred as PFTK model) can accurately predict TCP performance over a wide range of loss rates. We use this model as a basis of our work.

10ICCCN 2003

Possible application of the model

Fundamental trade off between the detection rapidness of actual losses versus the risk of unnecessary retransmissions:

small RTO: fast detection, more risk of spurious timeout; large RTO: slow detection, less risky but long stall time.

RTOmin has significant impact on RTO value, common practice is set it to 2*clock. This model can assist in determining an appropriate value of RTOmin since it considers spurious timeouts explicitly.

Help evaluating the impact of lower layer protocols’ settings on the performance of TCP.

Help evaluating different TCP modifications designed for alleviating the effect of Spurious Timeout.

11ICCCN 2003

Modelling Assumptions

The sending rate is not limited by the advertised receiver window, and the sender always has sufficient data to send.

Segment losses in a round are independent from losses in other rounds. All other segments which were sent after the first lost segment in a specific round are also lost.

The time required to send a window of data is smaller than an RTT.

No RTT fluctuation measurements caused by queuing delays. In the absence of delay spikes, the RTT measurements compose a stationary random process.

12ICCCN 2003

Stochastic model of long delay pattern

Variation of RTT showing delay spikes

Markov Chain model

13ICCCN 2003

One Long Delay Period (LDP)

14ICCCN 2003

One Long Delay Cycle (LDC)

15ICCCN 2003

Steady-state Sending Rate Estimation

Consider LDC as the basis for steady state sending rate calculation, instead of using NP in PFTK model.

The proposed model considers a larger time scale than PFTK model: one LDC is composed of several NP and one LDP.

A high-level expression of the model:

LDCNPmLDPLDPZ

NPNPZ

LDCG

LDPZENPZmEGEDIpB

one in ofnumber : one duringsent segments :

one duringsent segments :

duringsent segments total:)()(

)(),,(

16ICCCN 2003

Steady-state Throughput Estimation

Use the sending rate as the basis of throughput estimation.

Subtract dropped segments and multiple retransmitted segments for the same segment from total number of segments sent during NP.

Subtract dropped segments and spuriously retransmitted segments from total number of segments sent during LDP.

The duration of LDC unchanged for throughput estimation.

)()()

,(),,(

LDPZENPZmEGEDIpT

17ICCCN 2003

Simulation Setup

Topology:

Parameters:

18ICCCN 2003

Sending rate estimation comparison (200ms RTT)

RTT=200ms

E(I)=30s

RTT=200ms

E(I)=240s

19ICCCN 2003

Sending rate estimation comparison (400ms RTT)

RTT=400ms

E(I)=30s

RTT=400ms

E(I)=240s

20ICCCN 2003

Throughput estimation comparison (200ms RTT)

RTT=200ms

E(I)=30s

RTT=200ms

E(I)=240s

21ICCCN 2003

Throughput estimation comparison (400ms RTT)

RTT=400ms

E(I)=30s

RTT=400ms

E(I)=240s

22ICCCN 2003

Estimation error vs. LDF*

Sending rate estimation error

Throughput estimation error

LDF = E(D)/E(I)

23ICCCN 2003

Estimation error vs. RTT

Sending rate estimation error

Throughput estimation error

24ICCCN 2003

Estimation error vs. packet error rate

Sending rate estimation error

Throughput estimation error

25ICCCN 2003

Conclusion

The proposed model can characterize the impact of delay spikes with different patterns on TCP’s performance.

The proposed model is more accurate than the PFTK model in estimating the steady state sending rate and throughput of TCP, especially in presence of frequent long delays.

Future research can be made on applying the model in TCP RTO setting selection or lower layer retransmission protocol design evaluation.