Augmenting Mobile 3G Using WiFi
Aruna Balasubramanian, Ratul Mahajan, Arun Venkataramani
2011-04-04
Jimin Lee
jmlee@mmlab.snu.ac.kr
Multimedia and Mobile communications Laboratory
Outline

Introduction

Measurement

Wiffler
 Prediction-based offloading
 Fast switching

Evaluation

Conclusion
2/23
Introduction

Mobile Internet access is suffering today
 The ubiquitous deployment of cellular data networks has
drawn millions of users
 Mobile data is growing exponentially
 This is creating immense pressure on the limited spectrum of
networks
 Is more spectrum the answer?
3/23
Measurement

To study 3G and WiFi network characteristics
 What is the availability of 3G and WiFi networks as seen by a vehicle user?
 What are the performance characteristics of the two networks?

Testbeds
 Outdoor testbeds that include effects present in real vehicular
settings such as noise, interference, traffic patterns
 Conducted across three cities

Amherst, Seattle, Sfo
 Vehicular nodes with 3G and WiFi radios

Amherst: 20 buses

Seattle: 1car

Sfo: 1car
4/23
Measurement

Methodology
 The vehicles visit many locations multiple times each day

Amherst : 12days, Seattle: 6days, Sfo: 3days
 The software on the vehicle includes the two programs

First program scans the 3G and WiFi channels simultaneously

Second program sends and receives data to a server
 Both server and vehicle log the characteristics of the data transfer
5/23
Measurement

Availability
 The server and the vehicle periodically send data to each other
over UDP
 An interface(3G or WiFi) is considered available if at least one
packet was received in the interval
 Availability is defined as the number of available 1-second intervals
divided by the total number of intervals
6/23
Measurement

Availability (cont’d)
 WiFi availability is lower than 3G
7/23
Measurement

Performance
 To measure the upstream and downstream UDP throughput
 The server and the vehicle send 1500-byte packets every 20ms.
 WiFi throughput is lower than 3G
8/23
Measurement

Summary
 The availability of WiFi is poorer than 3G
 WiFi throughput is also much lower than 3G throughput

Augmenting 3G using WiFi
 How can we reduce 3G usage by using WiFi?
 The simplest policy

To send data on WiFi when available and switch 3G when WiFi is
unavailable

First, Availability of WiFi can be low : 11%

Second, WiFi throughput is lower than 3G
9/23
Wiffler

Key techniques
 Leveraging delay tolerance
 Exploit
the delay tolerance of apps to increase data offloaded to
WiFi
 Fast switching
 For
apps with strict quality of service requirements
 Such
as VoIP and video stream
10/23
Leveraging delay tolerance

The simplest solution
 To wait until the delay tolerance threshold to transfer
data on WiFi when available
 It

may significantly increase the completion time
So, Wiffler uses the predictor to estimate offload
capability of WiFi network
11/23
Leveraging delay tolerance (cont’d)

Prediction-based offloading
 Transfer required: S bytes by D seconds
 D: earliest delay tolerance threshold among queued transfers
 W: predicted WiFi capacity over future D seconds
Parallel
Operating
if(WiFi is available)
Send data on WiFi
If(W < S and 3G is available)
Send data on 3G
12/23
Leveraging delay tolerance (cont’d)

WiFi throughput prediction
 We predict WiFi offload capacity

Based on an estimate of the average throughput offered by an AP and
a prediction of the number of APs that will be encountered
 AP meetings occur in bursts
 So, we can predict the number of AP encounters using a history-
based predictor

Future AP encounters depend on recent past

The mobile node keeps track of the last N Aps

By using this information, we can compute the
• (# of APs) * (capacity per AP)
13/23
Fast switching to 3G

Poor WiFi connectivity will hurt demanding apps
 Such as VoIP, video streaming

If WiFi is losing or delaying packets, we should
send them on 3G as soon as possible
 Link-layer retransmissions take much time
 Variable medium access delays
14/23
Fast switching to 3G (cont’d)

Motivation
 Waiting for WiFi link-layer retransmissions incurs delay
 Losses are bursty in the vehicular environment

The simple mechanism
 It sends the packet on 3G if the WiFi link-layer fails to deliver
the packet within a delay threshold
 It’s better to send time-sensitive packets on 3G rather than
waiting for likely more failures on WiFi
15/23
Evaluation

Deployment on 20 vehicular nodes

Simulations
16/23
Evaluation

Deployment on 20 vehicular nodes
 Prediction-based offloading
Data offloaded to WiFi
Prediction-based offloading
30%
WiFi when available
11%
Transfer size: 5MB, Delay tolerance: 60 secs,
Inter-transfer gap: random with mean 100 secs
17/23
Evaluation

Deployment on 20 vehicular nodes
 Fast switching to 3G
Time w/ good voice quality
Fast switching
68%
WiFi when available
42%
VoIP-like traffic: 20-byte packet every 20 ms
With standard MOS metric
18/23
Evaluation

Simulations
 To evaluate Wiffler’s prediction-based offloading and
fast switching from others
 Alternative strategies
 Impatient
: use WiFi when available
 Patient
: waits until the threshold
 Oracle
: perfect future knowledge
19/23
Evaluation

Wiffler increases data offloading to WiFi
20/23
Evaluation

Prediction reduces completion time
21/23
Evaluation

Fast switching improves performance of
demanding apps
22/23
Conclusion

Paper develops techniques to combine mutiple interfacees
with different costs and ubiquitousness
 3G is costly but more ubiquitous
 WiFi is cheaper but intermittently available

It overcomes WiFi’s poor availability by leveraging delay
tolerance of applications and a fast switching mechanism
23/23