TCP/IP Performance over Asymmetric Networks · TCP/IP Performance over Asymmetric Networks ... • Normalized bandwidth ratio k determines the behaviour of TCP • On average, only

DISLabDISLab ICE HUFSICE HUFS

TCP/IP Performance over Asymmetric Networks

2007. 5. 15발표자 : 장주연

2 page

DISLabDISLab ICE HUFS ICE HUFS

ContentsContents

Introduction

Type of Network Asymmetry

Impact of Asymmetry on TCP Performance

Improving TCP Performance over Asymmetric Networks

Experimental Evaluation of Performance Improvement Techniques

Further Reading

Summary

DISLabDISLab ICE HUFSICE HUFS

IntroductionIntroduction

4 page

DISLabDISLab ICE HUFS ICE HUFS

ObjectivesObjectives

Explain types of asymmetry that are present in today’s networks

Comprehend specific performance issues when TCP/IP traffic is transported over asymmetric networks

Learn techniques to address TCP performance problems in asymmetric environments

DISLabDISLab ICE HUFSICE HUFS

Type of Network AsymmetryType of Network Asymmetry

6 page

DISLabDISLab ICE HUFS ICE HUFS

What is Network Asymmetry?What is Network Asymmetry?

Network asymmetry refers to the situation where characteristics in the uplink are different than those in the downlink

Examples• Cable modem• ADSL• Satellite

7 page

DISLabDISLab ICE HUFS ICE HUFS

Types of Network AsymmetryTypes of Network Asymmetry

Bandwidth asymmetry

Media-access asymmetry

Loss rate asymmetry

8 page

DISLabDISLab ICE HUFS ICE HUFS

Bandwidth AsymmetryBandwidth Asymmetry

Forward and reverse bandwidth are significantly different

Typically downlink bandwidth is 10-1000 times the uplink bandwidth

Example: Direct PC has a 400Kbps downlink and a 56Kbps dialup uplink

9 page

DISLabDISLab ICE HUFS ICE HUFS

MediaMedia--Access AsymmetryAccess Asymmetry

Can occur when transmitter and receiver use shared medium (wired or wireless), and

Transmitter experiences larger (smaller) MAC delay than receiver

Can happen in both cellular and packet radio networks

10 page

DISLabDISLab ICE HUFS ICE HUFS

LossLoss--Rate AsymmetryRate Asymmetry

Packet loss probability in the uplink may be different than that of downlink

This can happen if one of the links is more congested than the other, for example

Loss-rate asymmetry can occur in any network, and it may be a transient phenomenon

DISLabDISLab ICE HUFSICE HUFS

Impact of Asymmetry on TCP PerformanceImpact of Asymmetry on TCP Performance

12 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of Bandwidth AsymmetryImpact of Bandwidth Asymmetry

Unidirectional data transfer• File download from a server• Normalized bandwidth ratio k determines the behaviour of TCP• On average, only 1 ACK gets through for every k packets sent

– Increase the chance of data packet loss– Infrequent ACKs result in slower growth of congestion window– Loss of ACKs could cause long idle periods

• Example– 10-Mbps downstream channel, 100-Kbps upstream channel

» raw bandwidth ratio is 100– 1000-byts data packets, 40-byte ACKs

» the ratio of packet sizes is 25– k is 100/25 = 4

13 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of Bandwidth AsymmetryImpact of Bandwidth Asymmetry ((cont.cont.))

Bidirectional data transfer• Exacerbate the problem due to bandwidth asymmetry

– Interaction between data packets of the upstream transfer and ACKs of the downstream transfer

• Figure 10.3– Measurements of the downstream connection operating over 10-Mbps wireless cable

modem network• Figure 10.4

– Measurements of the upstream connection operating over a 9.6-Kbps dialup line

FIGURE 10.3 FIGURE 10.4

14 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of MediaImpact of Media--Access AsymmetryAccess Asymmetry

A central base station suffers lower MAC overhead than distributed nodes

MAC overhead makes it expensive to transmit packets in one direction when there is an ongoing data transfer in the opposite direction

Packet radio network technology• Topology• Half-Duplex Radios• Media-Access• Error Control

15 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of MediaImpact of Media--Access AsymmetryAccess Asymmetry ((cont.cont.))

Packet radio network technology• Topology

– Figure 10.5 : Topology of a packet radio network

• Half-Duplex Radios– the transmission power drowns incoming receptions in the same frequency band– Moving from transmitting to receiving mode takes a nontrivial amount of time– TT R : transmit-to-receive turnaround time– TR T : receive-to-transmit turnaround time

16 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of MediaImpact of Media--Access AsymmetryAccess Asymmetry ((cont.cont.))

Packet radio network technology (cont.)• Media-Access

– the predominant reason for poor performance is the interactions between the media-access (MAC) protocol and TCP

– the MAC protocol for contention resolution is based on a polling scheme, similar to the RTS/CTS protocol in the IEEE 802.11 standard

back-off time

NACK_CTS

17 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of MediaImpact of Media--Access AsymmetryAccess Asymmetry ((cont.cont.))

Packet radio network technology (cont.)• Error Control

– The reliable link-layer protocol used in this network for error control is a simple frame-by-frame protocol with a window size of 1

– High per-packet overhead» The need for the communicating peers to first synchronize via the RTS/CTS protocol» The significant turnaround time for the radios

– These variable latencies and queueing of ACKs adversely affect smooth data flow– An optimization of the link-layer, error-control protocol that piggybacks link-layer ACKs

with data frames reduces TCP round-trip variations to some extent

18 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of MediaImpact of Media--Access Asymmetry (cont.)Access Asymmetry (cont.)

Fig. 10.6• the packet sequence trace of a

measured 200-KB TCP transfer over an unloaded wired path and one wireless hop in the packet radio network

Sender timeout으로 발생

각각 9~12s로 지속

Reason : 긴 round-trip time variation 때문에 retransmission timeout이 매우길게 계산됨

이상적인 상태

• TCP data tranfer의 round-trip time estimate(srtt)는 상대적으로 상수

• 낮은 표준편차(mdev)를 가짐• TCP retransmission timeout

– srtt+ 4*mdev

19 page

DISLabDISLab ICE HUFS ICE HUFS

Impact of MediaImpact of Media--Access Asymmetry (cont.)Access Asymmetry (cont.)

Fig. 10.7• Twenty round-trip time samples. The

samples have a mean of about 2.5s and a standard deviation of about 1.5s

Retransmission timeout : 약 10s

Link-layer의 optimization• TCP round-trip variation을 어떤 범위로축소

DISLabDISLab ICE HUFSICE HUFS

Improving TCP Performance over Asymmetric Improving TCP Performance over Asymmetric NetworksNetworks

21 page

DISLabDISLab ICE HUFS ICE HUFS

TCP Performance Enhancements over Asymmetric TCP Performance Enhancements over Asymmetric NetworksNetworks

Two key issues need to be addressed• Manage bandwidth usage on the uplink

– Reduce the number of ACKs

• Avoid adverse impact of infrequent ACKs

Solutions• Local link-layer solutions• End-to-end techniques

Uplink Bandwidth Management• TCP header compression• ACK filtering• ACK congestion control• ACKs-first scheduling

Handling infrequent ACKs• TCP sender adaptation• ACK reconstruction

22 page

DISLabDISLab ICE HUFS ICE HUFS

Uplink Bandwidth ManagementUplink Bandwidth Management

Can be realized by:• Control the degree of compression• Control the frequency• Control the scheduling of upstream ACKs

TCP Header Compression• For use over low-bandwidth links running SLIP/PPP• Reduce the size of ACKs on the slow uplink• Some problems remain:

– MAC overhead» Independent of packet size

– Adverse interaction with large upstream data packets» Bidirectional traffic

23 page

DISLabDISLab ICE HUFS ICE HUFS

Uplink Bandwidth ManagementUplink Bandwidth Management (cont.)(cont.)

ACK Filtering (AF)• TCP-aware link-layer technique• Reduce the number of TCP ACKs sent on upstream channel• Router maintains states for connections that have ACKs packets enqueued• Remove “redundant” ACKs packets

– Duplicate ACKs not removed– Selective ACKs not removed

ACK Congestion Control (ACC)• Operate on an end-to-end basis• Apply congestion control to ACK packets• Mimic TCP congestion control mechanism• Employ delayed ACK

– One ACK sent for every d data packets received– One ACK acknowledges several data packets

• Example: RED+ECN

24 page

DISLabDISLab ICE HUFS ICE HUFS

Uplink Bandwidth ManagementUplink Bandwidth Management (cont.)(cont.)

ACKs-First Scheduling• ACK packets may be delayed by data packets in a FIFO queue

• Separate ACK packets from data packets

• Give priority to ACKs– ACK packets are usually small (compared with data packets)– Minimal impacts in data packets

• Large data packet still causes delay– Segment large data packet before transmission

25 page

DISLabDISLab ICE HUFS ICE HUFS

Handling Infrequent Handling Infrequent ACKsACKs

Done either end-to-end or locally at the constrained uplink• TCP Sender Adaptation (SA)• ACK Reconstruction (AR)

TCP Sender Adaptation (SA)• End-to-end technique• The number of back-to-back packets can be sent is bounded• Take into account the amount of data (rather than number of packets) received• Mimic the effect of delayed ACK algorithm

26 page

DISLabDISLab ICE HUFS ICE HUFS

Handling Infrequent Handling Infrequent ACKsACKs (cont.)(cont.)

ACK Reconstruction (AR)• Local technique

• Reconstruct the ACK stream after it has traversed the upstream direction bottleneck link

• Enable implementation of AF or ACC with changes to TCP senders

• Deploy a soft-state agent called ACK reconstructor at the upstream end

• ACK threshold determines the spacing between interspersed ACKs at the output

• TCP senders can increase their cwnd at the right rate– Avoid burst behaviour

DISLabDISLab ICE HUFSICE HUFS

Experimental EvaluationExperimental Evaluation Of Performance Of Performance Improvement TechniquesImprovement Techniques

28 page

DISLabDISLab ICE HUFS ICE HUFS

Experimental Experimental with with Bandwidth AsymmetryBandwidth Asymmetry

TCP Reno enhanced with ACC, AF, SA and AR

AF/AR and AF/SA have the best performance• Table 10.1• 15%--21% increase in throughput

Degree of burstiness is significantly reduced

SA/AR is effective in overcoming the burstiness that results from a lossy ACK stream

Random drop is superior to drop-tail

Metric Reno ACC/SA AF/SA AF/AR AF Alone

Throughput (Mbps) 6.71 6.95 7.82 8.57 5.16

Average cwnd (pkts) 66.7 62 65.3 104.6 43.8

Average RTT (ms) 79 70 65 97 65

29 page

DISLabDISLab ICE HUFS ICE HUFS

Experimental Evaluation:Experimental Evaluation:MediaMedia--Access AsymmetryAccess Asymmetry

Protocols investigated: TCP Reno, Reno with ACC/SA and Reno with AF/SA

AF and ACC with SA yield better performance than Reno• Fig. 10.8

AF/SA outperforms ACC/SA

Improvement in throughput• 25% for 1 wireless hop• 41% for 3 wireless hops

DISLabDISLab ICE HUFSICE HUFS

SummarySummary

31 page

DISLabDISLab ICE HUFS ICE HUFS

SummarySummary

Network asymmetry• bandwidth• media-access• loss rate

Solutions for efficient TCP in a variety of asymmetric conditions