On Effectively Exploiting Multiple Wireless Interfaces in ...

On Effectively Exploiting Multiple Wireless Interfaces in Mobile Hosts

Cheng-Lin Tsao and Raghupathy SivakumarGeorgia Institute of Technology

ACM CoNEXT ‘09Dec 4, 2009Rome, Italy

2

Introduction

• Multi-homed wireless devices– Laptops equipped with Wi-Fi, WWAN, Bluetooth, WiMAX

– Smartphones equipped with Wi-Fi and 2.5G/3G

• BlackBerry, iPhone, Google Android

– Only one interface used at a time

• Priority-based switching

• What is the best approach to leverage the multiple interfaces available at a mobile device in terms of user performance?

3

EVDO802.11b/g

Desktop server

Smartphone

(Google Android)Laptop

Verizon

3G network

T-Mobile

3G network

Georgia Tech

Wi-Fi network

802.11g HSDPA

Ethernet

Ethernet

Network Nightmare

WAN emulator

Internet

AT&T DSL

Experimental Testbed

• Real-life heterogeneous wireless testbed– Laptop

• Atheros 802.11a/b/g PCMCIA

• Verizon USB727 EVDO

– Google Android G1

• Embedded 802.11g

• T-Mobile HSDPA

– Desktop server

• WAN emulator on Wi-Fi to mimic Internet

4

0

2

4

6

8

10

12

14

16

laptop (802.11b) laptop (802.11g) google android

Wireless interfaces

TC

P T

hro

ughput

(Mbps)

Wi-Fi

3G

Motivation

• Simple aggregation– Ex. pTCP [ICNP`02]

– Sum of available BW

– Marginal benefits

B1

B2

B1

B1+B2

B2

• Achieving performance better than the sum of the parts?

• Leveraging heterogeneity– Wi-Fi: high bandwidth

– 3G: high availability, allocated resources

simple aggr

5

Super-Aggregation

• Concept: transferring data over multiple cooperating interfaces w/ high-layer knowledge

• Design– Three generic principles

– TCP-specific throughput enhancement in Wi-Fi + 3G

– Extension to other protocols/wireless interfaces

• Realization– Client software changes that work w/ legacy servers

– Layer-3.5 TCP acceleration in Wi-Fi + 3G networks

– Prototyped and evaluated on laptop and smartphone

6

0

5

10

15

20

25

laptop (802.11b) laptop (802.11g) google android

Wireless interfaces

Thro

ughput

(Mbps)

TCP

UDP

biUDP

Principle 1: Selective Offloading

• Concept– Selectively offload certain portions of the transferred data to the low-bandwidth interface

• Relation to TCP: self-contention

– Between uplink ACK and downlink DATA

– PHY/MAC overheads

– Degrading throughput by 30%~70%

– Verified with bidir. UDP

• 1464B data and 32B ack

Impact from self-contention in Wi-Fi

7

Solution 1: Offloading-ACK

• Idea– Diverting uplink ACKs from Wi-Fi to 3G

• Challenges– Insufficient bandwidth on 3G

– Long RTT on 3G degrades overall TCP throughput

• Solution– Fractional offloading

– Opportunistic operation

8

Offloading-ACK Details

• Fractional offloading– Offloading sustainable fractions on 3G

– Discarding ACKs carrying no additional information

• Opportunistic operation– Offloading when RTT inflation has little impacts, like cwnd more than a threshold

– Heuristic: ssthresh value of TCP

• Default value: 20 mss

9

0

5

10

15

20

25

0 2 4 6 8 10Time (sec)

Thro

ughput (M

bps)

0

20

40

60

80

100

120

Cw

nd a

nd s

sthre

sh (

mss

)

throughput

cwnd

ssthresh

0

5

10

15

20

25

0 2 4 6 8 10Time (sec)

Thro

ughput (M

bps)

0

20

40

60

80

100

120

Cw

nd a

nd s

sthre

sh (

mss

)

throughput

cwnd

ssthresh

Principle 2: Proxying

• Concept– Use the low-bandwidth interface for critical controlwhen the high-bandwidth one is temporarily down

• Relation to TCP: blackoutsImpact from blackout

– Blackout: fading or handoff

• Vehicle net: up to 75 sec1

– Impacts

• RTO timeout

• Unnecessary idle

• Slow slow-start

blackout

1V. Bychkovsky, B. Hull, A. Miu, H. Balakrishnan, and S. Madden, “A measurement study of vehicular internet access using in situ wi-fi networks,” Mobicom `06.

10

Solution 2: Proxying-blackout-freeze

• Idea– Use the 3G link to notify the TCP sender about blackouts on Wi-Fi

• Challenges– Real-time blackout detection with low overhead

– Freezing a TCP connection during blackout

• Solution– Hybrid blackout detection

– Freezing TCP with flow control

11

Proxying-blackout-freeze Details

• Freezing TCP with flow control– Sending a zero-window advertisement on 3G to make TCP enter persist mode

– Resuming the TCP connection w/ the original flow window via 3G

• Hybrid blackout detection– Passive monitoring of received packets/beacons

– Active probing when no activity for a certain period

• ICMP probing if no packet for more than 200 ms

12

0

2

4

6

8

10

0.0% 0.1% 0.3% 1.0% 3.0%

Packet loss rate

Thro

ughput (M

bps)

Principle 3: Mirroring

• Concept– Intelligently mirror the certain portion of the transferred data on the low-bandwidth interface

• Relation to TCP: random lossesImpact from random wireless losses– Caused by interference,

fading, or long distance

– Interpreted by TCP as congestion

– 0.1% packet loss rate reduces TCP throughput by 49%

13

Solution 3: Mirroring-loss-fetching

• Idea– Hide random losses in the original connection and fetch the lost packets in the mirror connection

• Challenges– Decoupling TCP congestion control and reliability

– Mirroring the TCP connection on 3G

– Efficiently fetching lost packets on 3G

• Solution– Loss distinction

– Connection mirroring

– Fast fetching

14

Mirroring-loss-fetching Details

• Loss distinction– Receiving corrupted frames indicates random losses

• TCP connection mirroring– Replay messages exchanged in the original connection

– Offset TCP sequence numbers

– Verify identical data received in the mirror connection

• Selected and fast fetching– Proactively acknowledge unneeded packets

– Place a guard time before fetching the desired packets• 256 ms in prototype

15

Super-Aggregation Architecture

• Software Architecture– Client-only changes

– Layer-3.5 middleware

– Transparency to TCP & link layers

• Integrated operations– Offloading-ACK

– Proxying-blackout-freeze

– Mirroring-loss-fetchingData of downstream traffic

ACK of downstream traffic

Data of upstream traffic

ACK of upstream traffic

.

Offloader

ACK Marker

Intf Characterizer

TCP

Wi-Fi Interface 3G Interface

Mirroring

Manager

Loss Hider

Fast Fetcher

Blackout

Handler

Blackout

Detector

16

0

5

10

15

20

Offloading-ACK Proxying-

blackout-freeze

Mirroring-loss-

fetching

Thro

ughput

(Mbps)

default TCP

simple aggregation

Performance Evaluation

• Performance metric– Overall throughput of bulk data transfer

– Improvement over simple aggregation

• Improvement on Android– Offloading: 26%

– Proxying: 35%• Blackouts (2 sec every 20 sec)

– Mirroring: 52%• Random packet loss 0.3%

super-aggregation

17

60.00 60.01 60.02 60.03Time (sec)

def

ault T

CP su

per

-aggre

gat

ion

data

ack

Offloading-ACK Analysis

• Avoiding TCP self-contention in Wi-Fi– Packets captured with tcpdump

– Self-contention observed in default TCP

– Offloading-ACK utilizes the Wi-Fi downlink

60.00 60.01 60.02 60.03Time (sec)

def

ault T

CP su

per

-aggre

gat

ion

data

ack

18

0

5

10

15

20

25

0 2 4 6 8 10Time (sec)

Thro

ughput (M

bps)

0

20

40

60

80

100

120

Cw

nd a

nd s

sthre

sh (

mss

)

throughput

cwnd

ssthresh

0

5

10

15

20

25

0 2 4 6 8 10Time (sec)

Thro

ughput (M

bps)

0

20

40

60

80

100

120

Cw

nd a

nd s

sthre

sh (

mss

)

tputcwndssthresh

Proxying-blackout-freeze Analysis

• Minimizing the impact from blackouts– Same blackout period introduced

– Avoiding slow start

– Quick resumption after link recovery

– Maintaining cwnd and ssthresh

0

5

10

15

20

25

0 2 4 6 8 10Time (sec)

Thro

ughput (M

bps)

0

20

40

60

80

100

120

Cw

nd a

nd s

sthre

sh (

mss

)

tputcwndssthresh

Proxying-blackout-freeze

blackout

19

0

100

200

300

400

500

0.4 0.5 0.6 0.7 0.8 0.9 1.0

Time (sec)

Seq

uen

ce n

um

ber

normal fetching (data)normal fetching (ack)

Mirroring-loss-fetching Analysis

• Recovering lost packets efficiently– Recovering 393 packets out of 100k packets (0.3% loss on Wi-Fi)

– Fast fetching recovers lost segments 36x faster

– Guard time makes sure packet delivery on 3G

0

100

200

300

400

500

0.4 0.5 0.6 0.7 0.8 0.9 1.0

Time (sec)

Seq

uen

ce n

um

ber fast fetching (data)fast fetching (ack)normal fetching (data)normal fetching (ack)guard

20

0

5

10

15

20

25

30

0 10 20 30 40 50 60 70 80 90 100Time (sec)

Thro

ughput (M

bps)

default TCPsimple-aggregation0

5

10

15

20

25

30

0 10 20 30 40 50 60 70 80 90 100Time (sec)

Thro

ughput (M

bps)

default TCPsimple-aggregationsuper-aggregation

Integrated Operation Performance

• Integrated operations– Evaluated with a scenario with blackouts and random losses

• Enter 3G when t=5

• 2-sec blackouts when t=50 & t=62.5

• 1% packet loss at AP3

– Improving throughput by 169% in the scenario

21

Related Works

• Simple aggregation– pTCP [ICNP`02], WAMP [Globecom`99], RMTP [ICNP`01], MC2

[ToMC`07], MAR [MobiSys`04]

– Wireless specific: R2CP [MobiCom`05] and PRISM [ToMC`07]

– Requiring two-point deployment

• TCP enhancement over a single wireless network– Random losses: Snoop [MobiCom`95], WTCP [MobiCom`99]

– Blackout: Freeze-TCP [Infocom`00]

• Multi-interface mechanisms for energy efficiency– CoolSpots [MobiSys`06], Cell2Notify [MobiSys`07], Context-for-Wireless [MobiCom`06]

22

Concluding Remarks

• Study super-aggregation of heterogeneous wireless interfaces

• Propose super-aggregation principles– Offloading-ACK

– Proxying-blackout-freeze

– Mirroring-loss-fetching

– Generalization to rate-adaptive video streaming and more other wireless technologies

• Design and prototype the integrated architecture

• Evaluate on laptop/smartphone in testbed

Thank you!