Download the full version of the textbook now at textbookfull.

com

Passive and Active Measurement 15th

International Conference PAM 2014 Los Angeles
CA USA March 10 11 2014 Proceedings 1st
Edition Michalis Faloutsos
https://textbookfull.com/product/passive-and-
active-measurement-15th-international-conference-
pam-2014-los-angeles-ca-usa-
march-10-11-2014-proceedings-1st-edition-michalis-
faloutsos/

Explore and download more textbook at https://textbookfull.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Runtime Verification 20th International Conference RV 2020

Los Angeles CA USA October 6 9 2020 Proceedings Jyotirmoy
Deshmukh
https://textbookfull.com/product/runtime-verification-20th-
international-conference-rv-2020-los-angeles-ca-usa-
october-6-9-2020-proceedings-jyotirmoy-deshmukh/
textbookfull.com

Smart Health International Conference ICSH 2014 Beijing

China July 10 11 2014 Proceedings 1st Edition Xiaolong
Zheng
https://textbookfull.com/product/smart-health-international-
conference-icsh-2014-beijing-china-july-10-11-2014-proceedings-1st-
edition-xiaolong-zheng/
textbookfull.com

Reversible Computation 6th International Conference RC

2014 Kyoto Japan July 10 11 2014 Proceedings 1st Edition
Shigeru Yamashita
https://textbookfull.com/product/reversible-computation-6th-
international-conference-rc-2014-kyoto-japan-
july-10-11-2014-proceedings-1st-edition-shigeru-yamashita/
textbookfull.com

Women in the International Film Industry: Policy, Practice

and Power Susan Liddy

https://textbookfull.com/product/women-in-the-international-film-
industry-policy-practice-and-power-susan-liddy/

textbookfull.com
Vibrations in the Production System: Measurement and
Analysis with Water Jet Technology Štefánia Olejárová

https://textbookfull.com/product/vibrations-in-the-production-system-
measurement-and-analysis-with-water-jet-technology-stefania-olejarova/

textbookfull.com

Study of Electronic Properties of 122 Iron Pnictide

Through Structural Carrier Doping and Impurity Scattering
Effects 1st Edition Tatsuya Kobayashi (Auth.)
https://textbookfull.com/product/study-of-electronic-properties-
of-122-iron-pnictide-through-structural-carrier-doping-and-impurity-
scattering-effects-1st-edition-tatsuya-kobayashi-auth/
textbookfull.com

Complete Children s Cookbook Delicious Step by Step

Recipes for Young Cooks 2nd Edition Dk

https://textbookfull.com/product/complete-children-s-cookbook-
delicious-step-by-step-recipes-for-young-cooks-2nd-edition-dk/

textbookfull.com

Heroes and Heroism in British Fiction Since 1800: Case

Studies 1st Edition Barbara Korte

https://textbookfull.com/product/heroes-and-heroism-in-british-
fiction-since-1800-case-studies-1st-edition-barbara-korte/

textbookfull.com

The Oxford Edition of Blackstone’s: Commentaries on the

Laws of England: Book I: Of the Rights of Persons William
Blackstone
https://textbookfull.com/product/the-oxford-edition-of-blackstones-
commentaries-on-the-laws-of-england-book-i-of-the-rights-of-persons-
william-blackstone/
textbookfull.com
Building Organizational Intelligence A Knowledge
Management Primer 1st Edition Liebowitz

https://textbookfull.com/product/building-organizational-intelligence-
a-knowledge-management-primer-1st-edition-liebowitz/

textbookfull.com
 Michalis Faloutsos
Aleksandar Kuzmanovic (Eds.)
LNCS 8362

Passive and Active

Measurement
15th International Conference, PAM 2014
Los Angeles, CA, USA, March 10-11, 2014
Proceedings

123
Lecture Notes in Computer Science 8362
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Alfred Kobsa
University of California, Irvine, CA, USA
Friedemann Mattern
ETH Zurich, Switzerland
John C. Mitchell
Stanford University, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
Oscar Nierstrasz
University of Bern, Switzerland
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Germany
Madhu Sudan
Microsoft Research, Cambridge, MA, USA
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbruecken, Germany
Michalis Faloutsos
Aleksandar Kuzmanovic (Eds.)

Passive and Active

Measurement
15th International Conference, PAM 2014
Los Angeles, CA, USA, March 10-11, 2014
Proceedings

13
Volume Editors
Michalis Faloutsos
University of New Mexico
Computer Science Department
Engineering Building II, Albuquerque, NM 87131, USA
E-mail: [email protected]
Aleksandar Kuzmanovic
Northwestern University
EECS Department
2145 Sheridan Road, Evanston, IL 60208, USA
E-mail: [email protected]

ISSN 0302-9743 e-ISSN 1611-3349

ISBN 978-3-319-04917-5 e-ISBN 978-3-319-04918-2
DOI 10.1007/978-3-319-04918-2
Springer Cham Heidelberg New York Dordrecht London

Library of Congress Control Number: 2014930886

LNCS Sublibrary: SL 5 – Computer Communication Networks

and Telecommunications
© Springer International Publishing Switzerland 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and
executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication
or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location,
in its current version, and permission for use must always be obtained from Springer. Permissions for use
may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution
under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication,
neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or
omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
 Preface

Welcome to the proceedings of the 2014 Passive and Active Measurement (PAM)
Conference. The event, which was held in Los Angeles this year, focused on
research in and the practice of Internet measurements. This was the 15th PAM.
Following its genesis in 2000, the conference has maintained a strong workshop
feel, providing an opportunity for the presentation of innovative and early work,
with lively discussion and active participation from attendees.
In 2012 the conference broadened its scope, reflecting the widening uses of
network measurement and analysis methods. The aim was to facilitate the un-
derstanding of the expanding role that measurement techniques play as they
become building blocks for a variety of networking environments, application
profiling, and cross-layer analysis. In 2014 we continued with this wider scope,
although we did not neglect PAM’s core topics.
PAM 2014 attracted 76 submissions. The papers came from academia and
industry from around the world. It was especially pleasing to see the global
nature of submissions.
The Technical Program Committee was chosen from a group of experts in
Internet measurement, drawing on past contributors to PAM including distin-
guished academic and industrial researchers, but also with a group of first-time
members. Additionally, we aimed to have a strong global representation on the
committee, and achieved this with members from around the world.
The final program of 24 papers was selected after each submission was care-
fully reviewed by at least three members of the Program Committee (PC), at
least one of whom rated themselves as knowledgeable with regard to the content
of the paper. We were delighted with the quality of reviews – they were careful,
insightful, and paid attention to detail. The reviews were followed by an exten-
sive discussion phase. PAM has traditionally avoided a large PC meeting and
the difficulties it creates for a global PC and instead uses on-line discussions.
This year, these were impressively robust: Reviewers provided more than 350
comments on papers, some almost as detailed as the reviews themselves. Most
of the final papers were then shepherded by PC members.
This year’s conference also continued the selection criteria related to repro-
ducible research, which was established in 2013. It is our belief that one of the
most pressing issues in the field of Internet measurement research is the fact that
many papers report on data sets that are never disclosed. Hence, PAM strongly
encourages the authors to publish their data sets.
In addition, the PC selected seven papers to appear as posters at the
conference, and these are included in this volume as extended abstracts. The
final program included papers on a wide range of measurement topics, and in-
cluded authors from 13 countries and five continents. Our most sincere thanks
go to the PC members for their diligence and care in reviewing, discussing, and
VI Preface

shepherding the papers that appear here, and to Marcel Flores for organizing
and maintaining the HotCRP site for us.
We are also most grateful to the Steering Committee, Jelena Mirkovic, who
was the local chair, and Jedidiah Crandall who served as the publicity chair. We
hope that you enjoy the papers in these proceedings.

March 2014 Aleksandar Kuzmanovic

Michalis Faloutsos
 Organization

Organizing Committee
Conference Chair
Michalis Faloutsos The University of New Mexico, USA

Program Chair
Aleksandar Kuzmanovic Northwestern University, USA

Local Chair
Jelena Mirkovic USC Information Sciences Institute

Publicity Chair
Jedidiah Crandall The University of New Mexico, USA

Steering Committee
Fabio Ricciato University of Salento, Italy
George Riley Georgia Institute of Technology, USA
Ian Graham Endace, New Zealand
Neil Spring University of Maryland, USA
Nevil Brownlee The University of Auckland, New Zealand
Nina Taft Technicolor Palo Alto Research Center, USA
Matthew Roughan University of Adelaide, Australia
Rocky K. C. Chang The Hong Kong Polytechnic University

Program Committee
Alan Mislove Northeastern University, USA
Alberto Dainotti CAIDA, USA
Arun Venkataramani UMass, USA
Bernhard Ager ETH Zurich, Switzerland
Bin Liu Tsinghua University, China
Bruce Maggs Akamai and Duke University, USA
Constantine Dovrolis Georgia Tech, USA
David Choffnes Northeastern University, USA
Dmitri Logiunov Texas A&M, USA
Fernando Silveira Technicolor, USA
Gabor Vattay Eotvos Larand University, Hungary
Han Song Narus Inc, USA
VIII Organization

Jelena Mirkovic ISI, USA

Marios Iliofotou Narus Inc., USA
Mark Allman ICSI, USA
Matthew Luckie CAIDA, USA
Michael Rabinovich Case Western Reserve University, USA
Minaxi Gupta Indiana University, USA
Myungjin Lee University of Edinburgh, UK
Neil Spring University of Maryland, USA
Paul Barford Wisconsin Madison, USA
Rade Stanojevic Telefonica, Spain
Sergey Gorinsky IMDEA, Spain
Thomas Karagiannis Microsoft, UK
Xenofontas Dimitropoulos ETH Zurich, Switzerland
Youngseok Lee CNU, Korea
Zhichun Li NEC Labs, USA
Zhi-Li Zhang University of Minnesota, USA

Sponsoring Institutions
University of New Mexico, USA
 Table of Contents

Internet Wireless and Mobility

RadioProphet: Intelligent Radio Resource Deallocation for Cellular
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Junxian Huang, Feng Qian, Z. Morley Mao, Subhabrata Sen, and
Oliver Spatscheck

Mobile Network Performance from User Devices: A Longitudinal,

Multidimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Ashkan Nikravesh, David R. Choffnes, Ethan Katz-Bassett,
Z. Morley Mao, and Matt Welsh

Diagnosing Path Inflation of Mobile Client Traffic . . . . . . . . . . . . . . . . . . . . 23

Kyriakos Zarifis, Tobias Flach, Srikanth Nori, David Choffnes,
Ramesh Govindan, Ethan Katz-Bassett, Z. Morley Mao, and
Matt Welsh

An End-to-End Measurement Study of Modern Cellular Data

Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Yin Xu, Zixiao Wang, Wai Kay Leong, and Ben Leong

Measurement Design, Experience and Analysis

A Second Look at Detecting Third-Party Addresses in Traceroute
Traces with the IP Timestamp Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Matthew Luckie and kc claffy

Ingress Point Spreading: A New Primitive for Adaptive Active Network

Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Guillermo Baltra, Robert Beverly, and Geoffrey G. Xie

On Searching for Patterns in Traceroute Responses . . . . . . . . . . . . . . . . . . . 67

Nevil Brownlee

Volume-Based Transit Pricing: Is 95 the Right Percentile? . . . . . . . . . . . . . 77

Vamseedhar Reddyvari Raja, Amogh Dhamdhere,
Alessandra Scicchitano, Srinivas Shakkottai, kc claffy, and
Simon Leinen
X Table of Contents

Performance Measurement
Dissecting Round Trip Time on the Slow Path with a Single Packet . . . . 88
Pietro Marchetta, Alessio Botta, Ethan Katz-Bassett, and
Antonio Pescapé

Is Our Ground-Truth for Traffic Classification Reliable? . . . . . . . . . . . . . . . 98

Valentı́n Carela-Español, Tomasz Bujlow, and Pere Barlet-Ros

Detecting Intentional Packet Drops on the Internet via TCP/IP Side

Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Roya Ensafi, Jeffrey Knockel, Geoffrey Alexander, and
Jedidiah R. Crandall

The Need for End-to-End Evaluation of Cloud Availability . . . . . . . . . . . . 119

Zi Hu, Liang Zhu, Calvin Ardi, Ethan Katz-Bassett,
Harsha V. Madhyastha, John Heidemann, and
Minlan Yu

Protocol And Application Behavior

Exposing Inconsistent Web Search Results with Bobble . . . . . . . . . . . . . . . 131
Xinyu Xing, Wei Meng, Dan Doozan, Nick Feamster,
Wenke Lee, and Alex C. Snoeren

Modern Application Layer Transmission Patterns from a Transport

Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Matt Sargent, Ethan Blanton, and Mark Allman

Third-Party Identity Management Usage on the Web . . . . . . . . . . . . . . . . . 151

Anna Vapen, Niklas Carlsson, Anirban Mahanti, and
Nahid Shahmehri

Understanding the Reachability of IPv6 Limited Visibility Prefixes . . . . . 163

Andra Lutu, Marcelo Bagnulo, Cristel Pelsser, and Olaf Maennel

Characterization of Network Behavior

Violation of Interdomain Routing Assumptions . . . . . . . . . . . . . . . . . . . . . . 173
Riad Mazloum, Marc-Olivier Buob, Jordan Augé, Bruno Baynat,
Dario Rossi, and Timur Friedman

Here Be Web Proxies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Nicholas Weaver, Christian Kreibich, Martin Dam, and Vern Paxson

Towards an Automated Investigation of the Impact of BGP Routing

Changes on Network Delay Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Massimo Rimondini, Claudio Squarcella, and Giuseppe Di Battista
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
 Table of Contents XI

Peering at the Internet’s Frontier: A First Look at ISP Interconnectivity

in Africa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Arpit Gupta, Matt Calder, Nick Feamster, Marshini Chetty,
Enrico Calandro, and Ethan Katz-Bassett

Network Security and Privacy

Assessing DNS Vulnerability to Record Injection . . . . . . . . . . . . . . . . . . . . . 214
Kyle Schomp, Tom Callahan, Michael Rabinovich, and Mark Allman
How Vulnerable Are Unprotected Machines on the Internet? . . . . . . . . . . . 224
Yuanyuan Grace Zeng, David Coffey, and John Viega
A Closer Look at Third-Party OSN Applications: Are They Leaking
Your Personal Information? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Abdelberi Chaabane, Yuan Ding, Ratan Dey,
Mohamed Ali Kaafar, and Keith W. Ross
On the Effectiveness of Traffic Analysis against Anonymity Networks
Using Flow Records . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Sambuddho Chakravarty, Marco V. Barbera, Georgios Portokalidis,
Michalis Polychronakis, and Angelos D. Keromytis

Poster Abstracts
Scaling Bandwidth Estimation to High Speed Networks . . . . . . . . . . . . . . . 258
Qianwen Yin, Jasleen Kaur, and F. Donelson Smith
Scalable Accurate Consolidation of Passively Measured Statistical
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Silvia Colabrese, Dario Rossi, and Marco Mellia
A Needle in the Haystack - Delay Based User Identification in Cellular
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Marco V. Barbera, Simone Bronzini, Alessandro Mei, and
Vasile C. Perta
Understanding HTTP Traffic and CDN Behavior from the Eyes
of a Mobile ISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Pedro Casas, Pierdomenico Fiadino, and Arian Bär
On Understanding User Interests through Heterogeneous Data
Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Samamon Khemmarat, Sabyasachi Saha, Han Hee Song,
Mario Baldi, and Lixin Gao
XII Table of Contents

Nightlights: Entropy-Based Metrics for Classifying Darkspace Traffic

Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Tanja Zseby, Nevil Brownlee, Alistair King, and kc claffy

Distributed Active Measurement of Internet Queuing Delays . . . . . . . . . . . 278

Pellegrino Casoria, Dario Rossi, Jordan Augé, Marc-Olivier Buob,
Timur Friedman, and Antonio Pescapé

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281

 RadioProphet: Intelligent Radio Resource Deallocation
for Cellular Networks

Junxian Huang1, Feng Qian2 , Z. Morley Mao3 ,

Subhabrata Sen2 , and Oliver Spatscheck2
1
Google Inc.
2
AT&T Labs – Research
3
University of Michigan

Abstract. Traditionally, radio resources are released in cellular networks by

statically configured inactivity timers, causing substantial resource inefficiencies.
We propose a novel system RadioProphet (RP), which dynamically and
intelligently determines in real time when to deallocate radio resources by
predicting the network idle time based on traffic history. We evaluate RP using 7-
month-long real-world cellular traces. Properly configured, RP correctly predicts
85.9% of idle time instances and achieves radio energy savings of 59.1% at the
cost of 91.0% of signaling overhead, outperforming existing proposals. We also
implement and evaluate RP on real Android devices, demonstrating its negligible
runtime overhead.

1 Introduction
Cellular networks employ a specific radio resource management policy distinguishing
them from wired and Wi-Fi networks. Previous studies [5][10][8] have shown that in
cellular networks, the origin of low resource efficiency comes from the way resources
are released. To avoid high signaling load, radio resources are only released after an idle
time (also known as the “tail time” or Ttail ) controlled by statically configured inactivity
timers. During the tail time, energy is essentially wasted by the radio interface.
Without knowing when network traffic will occur, long tail timer settings (e.g., 11.6
seconds configured by an LTE network [8]) are essentially a conservative way to ensure
low signaling overhead, which is known to be a bottleneck for cellular networks. Given
that application behaviors are not random, using a statically configured timer is clearly
suboptimal. A smaller static timer value helps reduce radio energy, but is not an option
due to the risk of overloading cellular networks caused by signaling load increase.
An attractive alternative is to configure the timer dynamically — adaptively per-
forming radio resource release signaled by the handset by monitoring the traffic and
accommodating different traffic patterns. But the key challenge is determining when
to release resources, which essentially comes down to accurate and efficient prediction
of the idle time period. Clearly, the best time to do so is when the handset is about to
experience a long idle time period, otherwise the incurred resource allocation overhead
(i.e., signaling load) might be unacceptably high. Therefore, accurate and efficient
prediction of the idle time period is a critical prerequisite for dynamic timer schemes.
This paper proposes RadioProphet (RP), a practical system running on a handset
that makes dynamic decisions to deallocate radio resources based on accurate and
efficient prediction of network idle times. It makes the following contributions.

M. Faloutsos and A. Kuzmanovic (Eds.): PAM 2014, LNCS 8362, pp. 1–11, 2014.

c Springer International Publishing Switzerland 2014
2 J. Huang et al.

First, RP utilizes standard online machine learning (ML) algorithms to accurately

predict the network idle time, and performs resource deallocation only when the idle
time is sufficiently long. We explored various ML algorithms and prediction models
with tunable parameters, with the main contribution of using a measurement-driven
approach to find robust and easy-to-measure features, whose complex interaction with
the network idle time can be automatically discovered by the ML algorithms. The model
is validated using seven-month-long traces collected from real users (§5).
Second, we implement RP on a real Android smartphone to demonstrate its negligible
energy and CPU overhead. In contrast, all previous proposals [10][4][7] only perform
trace-driven simulation. To reduce the runtime overhead, RP strategically performs
binary prediction (i.e., whether the idle time is short or long) at the granularity of
a traffic burst consisting of a packet train sent or received in a batch. Compared to
fine-grained prediction of the precise value of packet inter-arrival time, our proposed
approach is much more efficient while yielding similar optimization results.
Third, we overcome critical limitations of previously proposed approaches, i.e.,
RadioJockey [4] and MakeIdle / MakeActive [7] are only applicable to background
applications without user interaction, with the ideal usage scenario of RadioJockey for
a single application only. With multiple concurrent applications, it suffers from low
prediction accuracy with increased overhead. In contrast, RP is specifically designed for
both foreground and background traffic. Since its prediction is based on the aggregate
traffic of all apps, RP incurs no additional overhead for supporting concurrent apps.
Fourth, we conduct comprehensive measurement of RP using real-world smartphone
traces (7 months from 20 users). The overall prediction accuracy is 85.9%. RP achieves
radio energy saving by 59.1%, at the cost of 91.0% additional signaling overhead in LTE
networks, significantly outperforming previous proposals. To achieve the same energy
saving, the additional signaling overheads incurred by MakeIdle [7] and naı̈ve fast
dormancy [1] are 305% and 215%, respectively. The maximal energy saving achieved
by RadioJockey [4] is only 27% since it is only applicable to background traffic.
Paper Organization. We provide sufficient background in §2 before giving an overview
of the RadioProphet (RP) system in §3. We detail how we select relevant features for
idle time prediction in §4, and then systematically evaluate RP in §5. In §6, we describe
related work before concluding the paper.

2 Background
In cellular networks, there is a radio resource control (RRC) state machine that
determines radio resource usage based on application traffic patterns, affecting device
energy consumption and user experience. Conceptually similar RRC state machines
exist in different types of cellular networks from 2G to 4G LTE. In 3G UMTS networks,
there are usually three RRC states [11]: idle, low-power state, and high-power state. In
4G LTE networks, there are only two RRC states: idle and active [8]. Note that RP
works for any type of RRC state machine with fast dormancy (described soon) support.
State Transitions. There are two types of state transitions. State promotions switch
from a low-power state to a high-power state. They are triggered by user data transmis-
sion in either direction. State demotions go in the reverse direction, usually triggered
 RadioProphet: Intelligent Radio Resource Deallocation for Cellular Networks 3

by inactivity timers configured by the radio access network (RAN). For example, for a
commercial LTE network [8], at the active state, the RAN resets the timer to a constant
threshold Ttail =11.6 seconds whenever it observes any data frame. If there is no user
data transmission for Ttail seconds, the timer expires and the state is demoted to idle.
Similar timers exist in 3G networks (e.g., 12 seconds [11]).
State promotions incur long “ramp-up” delays of up to several seconds during which
tens of control messages are exchanged between the handset and the RAN for resource
allocation. Excessive state promotions increase the signaling overhead at the RAN and
degrade user experience, especially for short data transfers [3][10]. On the other hand,
state demotions incur tail times (Ttail ) causing waste of radio resources and handset
energy [5]. During the tail time, no data is transferred but the handset radio power is
much higher than that at the idle state (e.g., 1060mW vs 11mW for LTE [8]).
Fast Dormancy. Why are tail times necessary? First, the overhead of resource allo-
cation (i.e., state promotions) is high and tail times prevent frequent allocation and
deallocation of radio resources. Second, the RAN has no easy way of predicting the
network idle time of a handset, so it conservatively appends a tail to every network
usage period. This naturally gives rise to the idea of letting the handset actively
request for immediate resource release. Based on this intuition, a feature called Fast
Dormancy has been included in 3GPP since Release 7 [1][2]. It allows a handset
to send a control message to the RAN to immediately demote the RRC state to
idle (or a hibernating state) without experiencing the tail time. Fast dormancy is
supported by many handsets [2]. It can dramatically reduce the radio resource and the
handset energy usage with the potential penalty of increased signaling load when used
aggressively [3][10].

3 The RadioProphet (RP) System

The static tail times are the root cause of low resource efficiency in cellular networks.
RP leverages fast dormancy to dynamically determine when to release radio resources.
Challenge 1: trading off between resource saving and signaling load. The best
time to perform resource deallocation is when the handset is about to experience a long
idle time period t. If t is longer than the tail time, deallocating resources immediately
saves resources without any penalty of signaling load (i.e., state promotions). Other-
wise, doing so incurs an additional state promotion. Balancing such a critical tradeoff
requires predicting the idle time between data transfers so that fast dormancy is only
invoked when the idle time is sufficiently long.
Challenge 2: handling both foreground and background traffic. Idle time pre-
diction is particularly difficult for applications involving user interactions. Previous
systems, such as RadioJockey [4] and MakeActive [7], simply avoid this by only
handling traffic generated by applications running in the background.
Challenge 3: trading off between prediction accuracy and system performance.
RP is a service running on a handset with limited computational capabilities and
more importantly, limited battery life. So we need to minimize the overhead without
sacrificing much of the prediction accuracy.
4 J. Huang et al.

To address Challenge 1, we establish a novel machine-learning-based framework

for idle time prediction. Besides measuring the effectiveness and efficiency of a wide-
range of ML algorithms, our key contribution is addressing the hard problem of
selecting discriminating features that are relevant to idle time prediction. Based on
extensive measurement, we find that strategically using a few simple features (e.g.,
packet direction and size) leads to high prediction accuracy (§4). To address Challenge
2, we designed a general prediction framework that works for the aggregated (possibly
concurrent) traffic containing both foreground and background traffic. In contrast,
previous systems such as RadioJockey have the ideal usage case for a single app.
Further, we leverage the screen status [9], which indicates whether a user is interacting
with the device, to customize the prediction for screen-on and off traffic. Such a novel
approach can better balance the aforementioned tradeoff between resource saving and
signaling load. To address Challenge 3, RP performs binary prediction at the granularity
of a traffic burst consisting of a train of packets. In other words, we find that the
knowledge of whether the inter-burst time (IBT) is short or long (determined by a
threshold) is already accurate enough for guiding the resource deallocation. Such an
approach is much more efficient while yielding similar accuracy compared to the
expensive approach of predicting the precise value of packet inter-arrival time.
RP consists of three components: a traffic monitor, an IBT prediction module, and
a Fast Dormancy (FD) scheduler. The monitor inspects network traffic (only examines
packet headers) and extracts lightweight features for each burst in an online manner. The
features are then fed into the IBT prediction module, which trains models to predict the
IBT for the current burst. Then, the FD scheduler makes decision on whether to invoke
fast dormancy based on the IBT prediction result.
For IBT prediction, we formulate the traffic pattern as follows. The traffic is a
sequence of packets {Pi }(1 ≤ i ≤ n) in both directions. Let the timestamp of Pi be ti .
Using a burst threshold BT, the packets are grouped into bursts, i.e., {Pp , Pp+1 , ..., Pq }
belongs to a burst B if and only if: (1) tk+1 − tk ≤ BT for ∀k ∈ {p, ..., q − 1}, (2)
tq+1 − tq > BT, and (3) tp − tp−1 > BT. We define the inter-burst time IBT of burst
B to be the time gap following this burst, i.e., tq+1 − tq . We use a short IBT threshold
called SBT to classify an IBT, i.e., if IBT ≤ SBT, the burst is short, otherwise, it is long.
The IBT prediction module trains a model based on historical traffic information,
which consists of an array of bursts {B1 , ..., Bm }. Each Bi is a vector (f1 , f2 , ..., ft ,
ibti ) where {f1 , ..., ft } is the list of features of Bi and ibti is the IBT following burst
Bi observed by the traffic monitor. Whenever there is an idle time of BT, i.e., a new bust
appears, the prediction process starts. The feature vector of the current burst {f1 , ..., ft }
is generated and fed to the prediction module, which predicts whether the IBT following
the current burst is short or long. If short, no change is made and the handset stays in the
tail, since a packet is likely to appear soon. Otherwise, the FD scheduler invokes fast
dormancy to save energy. The prediction model is customized for each handset, and is
dynamically updated to adapt to the recent traffic pattern.

4 Feature Selection
We describe the measurement dataset before studying the feature selection in §4.2.
 RadioProphet: Intelligent Radio Resource Deallocation for Cellular Networks 5

1 35 1
30 Any bursts
0.8 Short bursts 0.8

% of all bursts
25
0.6 0.6

CDF
20
CDF

Port 80
0.4 Port 443 15 0.4 Facebook
Port 5222 10 Google services framework
0.2 Port 5228 0.2 LiveProfile
5 Yahoo! Sportacular
Port 53
0 0 0
0 5 10 15 20 25 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60
Inter-burst time (sec) Packet length (bytes) Inter-burst time (sec)

Fig. 1. IBT distributions of Fig. 2. Distributions of bursts Fig. 3. IBT distributions of

bursts whose last packets have grouped by packet length of bursts whose last packets are
specific port numbers the last packet associated with specific apps

4.1 The UMICH Dataset

The measurement data used in this study, which we call the UMICH dataset, is collected
from 20 students at University of Michigan for seven months. The students were given
Motorola Atrix (11 of them) or Samsung Galaxy S smartphones (9 of them) running
Android. Our custom data collection software continuously runs in the background and
collects three types of data. (1) Packet traces (only headers are used in this study).
(2) The process name responsible for sending or receiving each packet. (3) Other
system information such as screen status. Over the seven months (May to Dec 2011) we
collected 152 GB data. Although both cellular and Wi-Fi traces were collected, in this
study, we only use cellular traces, which contribute to 57.8% of the total traffic volume.

4.2 Measurement Driven Feature Selection for Burst Classification

We use a measurement-driven approach to derive features for the prediction model by
analyzing the correlation between various features and the IBT. First, to predict whether
an IBT is short or long, we look at the burst right before the IBT, since we observe that
the correlations between the IBT and earlier bursts’ features are much weaker. Second,
the features are extracted from the last three packets of a burst. This is because in most
cases, bursts are small (53% of bursts consist of no more than 3 packets), and even for
large bursts, we can usually tell their nature based on the last three packets, e.g., TCP
three-way handshake. Third, we only inspect packet headers since examining payload
incurs much higher overhead and also because traffic is increasingly being encrypted.
The lightweight features of the last three packets1 used by RP are listed below:
(1) packet direction, (2) server port number, (3) packet length (including header), (4)
protocol field in IP header, (5) TCP flags field in TCP header (0 if not TCP), and (6)
application name associated with the packet. These features are selected empirically so
that they are most relevant to IBT based on our measurement. We show three features
below as examples. We start our analysis with BT =1s and SBT =3s. Later we explore
how different BT and SBT settings affect our results in a quantitative manner (§5.4).
Port Number. Figure 1 shows IBT distributions of the top 5 ports ordered from top to
bottom in the legend, e.g., 80 is the most popular port, across all users. IBT distributions
1
If a burst contains less than three packets, all features for the missing packet(s) have a value of
0.
6 J. Huang et al.

of different ports clearly differ, especially for port 53, whose sudden jump at IBT = 5
seconds corresponds to the DNS retransmission timeout on Android. We also observe
clusters of IBT values for many other ports. For example, most bursts over port 5222
have a 20-second IBT corresponding to the keep-alive periodicity of Facebook.
Packet Length. Figure 2 plots the distributions of last packet lengths of bursts with
short IBT (IBT <SBT) and all bursts. Most bursts end with small packets, i.e., 84.59%
have their last packets ≤ 100 bytes, as a large packet is typically in the middle of a burst.
We observe high correlation for a few packet lengths values. For example, for 121 bytes,
93.04% bursts have short IBTs. The machine learning algorithms could automatically
discover these rules for prediction.
Applications. In Figure 3, the legend shows the sorted list of apps contributing the
largest amount of bursts with Facebook ranked at top 1. The differences in IBT values
are clear across apps. We also observe that for some apps, their periodic transfer
behaviors contribute to clusters of specific IBT values, e.g., Facebook and LiveProfile.
The application information can be very efficiently obtained (e.g., on Android [11]).

5 Implementation and Evaluation

5.1 Implementation
Trace-Driven Evaluation. We implement simulators of RP, MakeIdle [7], and Radio-
Jockey [4] on a desktop (3.16 GHz Xeon CPU with 16GB memory) using Matlab.
They work with an RRC state machine simulator (§5.2). We use them to evaluate the
accuracy and resource savings of RP under various configurations (§5.3, §5.4), as well
as to compare RP with other optimization techniques (§5.5), using the UMICH trace.
Implementation on Real Android Phone. We also implement the full RP system on
a Samsung Galaxy S3 phone running Android 4.0.4 to evaluate its running overhead
(§5.6). A modified TcpDump program is used as the traffic monitor. The IBT prediction
module is implemented as a native Android application running in the background.

5.2 Evaluation Methodology

We use three metrics to evaluate RP: prediction accuracy, saved radio energy, and
increased signaling load. The accuracy is defined as the number of bursts whose
immediate IBT (short or long) are correctly predicted divided by the total number of
bursts in the input trace. The radio energy, denoted as E, is the energy consumed by
the handset radio interface. It is one of the most significant components for the overall
energy usage of a handset, along with screen and CPU energy [11]. We build an RRC
state machine simulator, which takes as input a packet trace and employs the LTE
radio energy model derived in our previous work [8] to calculate E (using a UMTS
model [11] yields qualitatively similar results). The signaling load, denoted as S, is
quantified by the number of state promotions, each incurring a fixed number of signaling
messages [4]. S is also computed by the RRC state machine simulator.
Assume when a specific user trace is evaluated without any optimization performed
(no fast dormancy), E and S are calculated to be Ed and Sd , respectively. When RP
 RadioProphet: Intelligent Radio Resource Deallocation for Cellular Networks 7

Table 1. Impact of α, β on the prediction Table 2. Summary of prediction models

accuracy (PerUserDynamic model)

α= 100 500 1000 2000 5000 Name Description Accuracy

β=1 81.5% 83.7% 84.2% 82.4% 80.2% PerUser Use most recent α bursts of a user
β=2 80.1% 81.4% 82.9% 82.0% 80.0% 84.2%
Dynamic to predict next β bursts for that user
β=5 79.8% 80.9% 81.4% 81.0% 79.3%
PerUser Use a fixed set of n bursts of a user
β = 10 79.4% 80.0% 80.9% 80.0% 79.0% 80.8%
Static to train a fixed model for that user
β = 20 78.9% 79.6% 80.2% 79.5% 78.7%
AllUser Use a fixed set of k bursts of all users
77.5%
Static train a fixed model for all users

is used, the resulting E and S become E  and S  , respectively. We define Δ(E) =

(Ed − E  )/Ed and Δ(S) = (S  − Sd )/Sd (usually both are positive). They correspond
to the reduction of the radio energy and the increase of the signaling load brought by
RP, respectively. RP’s goal is to maximize Δ(E) while minimizing Δ(S).

5.3 Prediction Model Comparison

In RP, we use recent traffic information of a user to train a model, denoted as
PerUserDynamic. Specifically, for each user, the most recent α bursts are used to
predict the next β bursts. We study the impact of α, β in Table 1, using the Ensemble
Bagging [6] learning algorithm as an example (number of trees set to 20). If α is too
small, there is not enough training data for learning; if α is too large, the user is more
likely to switch to new applications that generate different traffic patterns so previously
learned rules may not be useful. Based on Table 1, we choose α = 1000 and β = 1 that
maximize the accuracy. In practice, α and β could also be dynamically adjusted.
Table 2 compares the PerUserDynamic model with two other models,
PerUserStatic (a fixed model for each user) and AllUserStatic (a fixed model for
all 20 users). For fair comparison, we use the same ML algorithm (Ensemble Bagging)
as used in Table 1. We set α = 1000 and β = 1 for the PerUserDynamic model as
discussed previously, and use n = 10,000 for PerUserStatic and k = 10,000 for
AllUserStatic (n and k defined in Table 2). Similar to Table 1, n and k are empiri-
cally selected to yield good prediction accuracies. We observe that PerUserDynamic
has higher prediction accuracy than the other two models, suggesting that it is necessary
to have a dynamic model for each user whose traffic pattern may be different from
others.

5.4 Selecting Burst Thresholds

We study the impact of BT and SBT (previous evaluations use BT =1s and SBT =3s). In
Table 3, S0 to S4 correspond to representative (BT, SBT) pairs. We find that aggressively
using a short SBT (S1 ) can significantly increase Δ(S). Among all settings, S4 yields
the highest Δ(E)/(1 + Δ(S)) value (the average radio energy saving per unit of
signaling load). It quantifies how well the balance between Δ(E) and Δ(S) is handled.
As mentioned in §3, configuring screen-on and off settings differently may yield
better optimization results, as screen-off traffic is usually generated by background apps
without user interaction, leading to statistically longer IBT. Therefore a more aggressive
8 J. Huang et al.

Table 3. Impact of BT and IBT (Classification Tree Table 4. Performance and accuracy of
with PerUserDynamic model, α=1000, β=1) different ML algorithms
Δ(E)
Settings (unit: sec) Accuracy Δ(E) Δ(S) (1+Δ(S))
ML Prediction time
Accuracy
S0 BT: 1 SBT: 3 82.65% 52.10% 101.64% 0.26 Algorithm (Training time)
S1 BT: 1 SBT: 2 84.80% 56.69% 158.99% 0.22 2.5 ms
Naı̈ve Bayes 76.1%
S2 BT: 1 SBT: 4 81.94% 49.07% 83.34% 0.27 6.4 ms
S3 BT: 0.5 SBT: 3 84.71% 53.74% 100.36% 0.27 Classification 5.9 ms
85.9%
S4 BT: 1.5 SBT: 3 85.39% 58.85% 93.75% 0.30 Tree 136.9 ms
S5 BT: 1/1.5 off/on Ensemble 106.6 ms
85.88% 59.07% 91.01% 0.31 87.4%
SBT: 2.5/3 off/on Bagging 626.1 ms

setting (smaller BT and SBT) can be applied to screen-off traffic without incurring much
signaling overhead. In Table 3, S5 is such a screen-aware setting. Compared with S4 ,
S5 saves more energy with less signaling overhead incurred. In fact, S5 achieves results
comparable to the optimal scenario to be shown in Table 5. This also indicates that
dynamically changing BT and SBT can help improve the effectiveness of RP.

5.5 Comparing Fast Dormancy Based Resource Optimization Approaches

Table 5 compares various optimization techniques using the UMICH dataset.

Basic fast dormancy. We set Ttail to a fixed value smaller than its original value.
RadioJockey [4] uses system calls to predict the end-of-session (EOS) for back-
ground app without user interaction, with the ideal usage scenario for a single app.
Given that we do not have system call traces in our dataset, we make two assumptions in
our simulation: (1) we use end-of-burst to approximate end-of-session, (2) RadioJockey
has high prediction accuracies (90% and 100%) for both single and concurrent apps
(although in reality, it performs worse when concurrent apps exist). A key limitation of
RadioJockey is it does not handle foreground traffic and only works when the screen is
idle (see §6), so we only apply RadioJockey to screen-off traffic2 .
MakeIdle [7] computes a wait time Twait that maximizes the energy saving if Ttail
is set to Twait for the previous M packets, it then applies this Twait for the next N
packets. The range we search for the optimal Twait is [0.5, 11.5] seconds, as suggested
by the authors. Since no recommendations have been made for the values of M and N ,
we empirically select different combinations of (M, N ) pairs.
RadioProphet : we explore three off-the-shelf machine learning algorithms with
the PerUserDynamic model (α=1000 and β=1): Naı̈ve Bayes, Classification Tree, and
Ensemble Bagging. Their performance and accuracy are summarized in Table 43 .
We now discuss the results in Table 5. “Fast dormancy 1s” is an aggressive approach
incurring unacceptable signaling overhead. “Fast dormancy 3s” reduces Δ(S) with
less energy saving as expected. For both approaches, their Δ(E)/(1 + Δ(S)) values
2
We configured short screen timeout for the 20 phones so screen-off is good approximation for
screen-idle.
3
The performance numbers in Table 4 correspond to the execution time of the scripts written
in Matlab on desktop. Our real implementation on the S3 smartphone uses C++ so it is much
more efficient (§5.6).
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
 RadioProphet: Intelligent Radio Resource Deallocation for Cellular Networks 9

Table 5. Comparison of optimization approaches. For RP, we use the PerUserDynamic model
(α=1000, β=1) with setting S5 in Table 3. RadioJockey is only applicable to screen-off traffic.

Δ(E)
Name Description & Configuration Δ(E) Δ(S) (1+Δ(S))

Basic Fast dormancy 1s Invoke fast dormancy after 1s idle time 62.7% 214.9% 0.20
Basic Fast dormancy 3s Invoke fast dormancy after 3s idle time 40.9% 95.8% 0.21
RadioJockey RadioJockey applied to 0.20
30.1% 51.7%
Assuming 100% accuracy only screen-off traffic (screen-off)
RadioJockey RadioJockey applied to 0.18
27.2% 52.0%
Assuming 90% accuracy only screen-off traffic (screen-off)
MakeIdle MakeIdle: based on previous M packets,
64.9% 305.2% 0.16
M:1000, N:100 predict next N packets
MakeIdle MakeIdle: based on previous M packets,
44.9% 195.2% 0.15
M:10, N:10 predict next N packets
Naı̈ve Bayes classification with mvmn:
RP: Naı̈ve Bayes 53.0% 107.9% 0.25
multivariate multinomial distribution
RP: Classification Tree Binary decision tree for classification 59.1% 91.0% 0.31
Method: Bag; type: classification
RP: Ensemble Bagging 59.3% 90.2% 0.31
weak leaner: decision tree; # of trees: 20
RP: Optimal Predict all IBTs correctly 59.8% 85.4% 0.32

(the average radio energy saving per unit of signaling load) are low due to a lack of
adaptation to dynamic traffic patterns.
For RadioJockey, by assuming the prediction accuracy for each background app to be
90%, it saves 27.2% of radio energy with 52% of signaling load, which can be slightly
improved when the accuracy increases to 100%. The overall saving is lower than that
of RP because RadioJockey does not handle foreground traffic usually triggered by
user interaction (§6). For MakeIdle, we use two representative (M, N ) settings. In both
cases, the incurred signaling load is prohibitive, since MakeIdle does not consider the
very important signaling load metric in its optimization framework.
For RP, in the optimal case assuming 100% prediction accuracy, it saves 59.8% of
radio energy with 85.4% of signaling load incurred. The signaling load is not zero,
because for IBTs smaller than Ttail but larger than SBT, even if the prediction is correct,
invoking fast dormancy would still incur an extra state promotion. This is inherent for
any fast dormancy based optimization technique. Among the three machine learning
algorithms, Ensemble Bagging achieves the best results, likely due to its usage of
multiple submodels to avoid overfitting. However, as shown in Table 4, its runtime
overhead is very high. The Classification Tree approach achieves similar optimization
results with much lower runtime overhead. The Δ(E)/(1 + Δ(S)) metric indicates that
RP outperforms other approaches in balancing Δ(E) and Δ(S).

5.6 Running Overhead on Real Phone

We implement the RadioProphet system on Android as discussed in §5.1, in order
to demonstrate its practicality on today’s smartphones. We breakdown its runtime
10 J. Huang et al.

overhead into three components: (1) traffic monitoring and feature extraction, (2) model
training and prediction, and (3) fast dormancy invocation. We found invoking fast
dormancy incurs negligible overhead. We therefore focus on (1) and (2) below.
Traffic Monitoring and Feature Selection. Unlike RadioJockey requiring system call
instrumentation, RP only needs to monitor packet traces, which is also needed by
RadioJockey. On the S3 smartphone, our traffic monitor incurs no more than 1% of
CPU overhead for parsing packet headers and generating burst features, although the
overhead is much lower when the throughput is low (e.g., less than 200 kbps). The
additional power to run the data collector is less than 17mW most of the time. In
contrast, the LTE radio power is at least 1000 mW [8].
Model Training and Prediction: Our implementation on S3 uses the Classification
Tree model that balances between accuracy and performance (Table 4). We measure
the average model training time to be 200ms and the average prediction time to be
0.1ms. Its incurred power overhead is always negligible (less than 10 mW).

6 Related Work and Concluding Remarks

We compare RP with three representative adaptive resource deallocation proposals.
TOP [10] leverages fast dormancy to eliminate the tail. It assumes each individual
application can predict an imminent long IBT with reasonable accuracy, and fast
dormancy is only invoked when the aggregate prediction across all concurrent apps
is long enough. TOP provides the prediction framework, but it does not solve the
challenging prediction problem itself, which is the key focus of RP.
MakeIdle [7] uses packet timing to calculate the optimal idle time before invoking
fast dormancy, in order to maximize the radio energy saving. However, MakeIdle
considers minimizing radio energy as the only objective, leading to unacceptably high
signaling overhead shown in Table 5. It leaves the job of reducing the signaling load to
another algorithm called MakeActive [7] that changes the traffic pattern by shifting
packets. MakeActive does not work with foreground traffic that is usually delay-
sensitive, and even for background traffic, there is no guarantee that it does not affect
user experience. In contrast, RP does not rely on changing traffic patterns and it works
with both foreground and background traffic. It can in fact coexist with traffic shaping
based optimization techniques such as MakeActive and TailEnder [5].
RadioJockey [4] uses program execution traces to predict the end of communication
spurts and invoke fast dormancy when necessary. It however has several limitations. (1)
It needs heavy instrumentation i.e., requiring complete system call traces in addition to
packet traces, while RP only examines packet header information. (2) RadioJockey only
works for background app without user interaction, since “predicting EOS events for
foreground applications turns out to be challenging since user interactions can trigger
network communications at any point in time” [4]. (3) RadioJockey treats different
apps separately and does not predict start-of-session, hence when concurrent apps
exist, the prediction accuracy would be affected. In contrast, RP introduces a general,
lightweight, and effective framework that naturally optimizes concurrent traffic from
both foreground and background apps. RP achieves even better optimization results for
all traffic than RadioJockey does for only background traffic (Table 5).
 RadioProphet: Intelligent Radio Resource Deallocation for Cellular Networks 11

To conclude, we propose a novel, practical, and effective system called

RadioProphet that intelligently predicts long idle period using off-the-shelf machine
learning algorithms, and deallocate resources based on IBT prediction for cellular net-
works. Using 7-month data collected from 20 real users, we show that RP outperforms
existing proposals in balancing the key tradeoff between resource saving and signaling
load. We present the first implementation of adaptive resource deallocation using fast
dormancy, demonstrating the feasibility of RP on real smartphones. We believe RP is an
important step towards application-aware energy and resource optimization in wireless
networks.

Acknowledgements. This research was supported in part by the National Science

Foundation under grants CNS-1039657, CNS-1059372 and CNS-0964545.

References
1. UE “Fast Dormancy” behavior. 3GPP discussion and decision notes R2-075251 (2007)
2. Configuration of fast dormancy in release 8. 3GPP discussion notes RP-090960 (2009)
3. System Impact of Poor Proprietary Fast Dormancy. 3GPP discussion and decision notes RP-
090941 (2009)
4. Athivarapu, P., Bhagwan, R., Guha, S., Navda, V., Ramjee, R., Arora, D., Padmanabhan, V.,
Varghese, G.: RadioJockey: Mining Program Execution to Optimize Cellular Radio Usage.
In: MobiCom (2012)
5. Balasubramanian, N., Balasubramanian, A., Venkataramani, A.: Energy Consumption in
Mobile Phones: A Measurement Study and Implications for Network Applications. In: IMC
(2009)
6. Breiman, L.: Bagging Predictor. Machine Learning 24(2) (1996)
7. Deng, S., Balakrishnan, H.: Traffic-Aware Techniques to Reduce 3G/LTE Wireless Energy
Consumption. In: CoNEXT (2012)
8. Huang, J., Qian, F., Gerber, A., Mao, Z.M., Sen, S., Spatscheck, O.: A Close Examination of
Performance and Power Characteristics of 4G LTE Networks. In: MobiSys (2012)
9. Huang, J., Qian, F., Mao, Z.M., Sen, S., Spatscheck, O.: Screen-Off Traffic Characterization
and Optimization in 3G/4G Networks. In: Proc. ACM SIGCOMM IMC (2012)
10. Qian, F., Wang, Z., Gerber, A., Mao, Z.M., Sen, S., Spatscheck, O.: TOP: Tail Optimization
Protocol for Cellular Radio Resource Allocation. In: Proc. ICNP (2010)
11. Qian, F., Wang, Z., Gerber, A., Mao, Z.M., Sen, S., Spatscheck, O.: Profiling Resource Usage
for Mobile Applications: a Cross-layer Approach. In: MobiSys (2011)
 Mobile Network Performance from User Devices:
A Longitudinal, Multidimensional Analysis

Ashkan Nikravesh1 , David R. Choffnes2 , Ethan Katz-Bassett3 ,

Z. Morley Mao1 , and Matt Welsh4
1
University of Michigan
2
Northeastern University
3
University of Southern California
4
Google Inc.

Abstract. In the cellular environment, operators, researchers and end users have
poor visibility into network performance for devices. Improving visibility is chal-
lenging because this performance depends factors that include carrier, access
technology, signal strength, geographic location and time. Addressing this re-
quires longitudinal, continuous and large-scale measurements from a diverse set
of mobile devices and networks.
This paper takes a first look at cellular network performance from this per-
spective, using 17 months of data collected from devices located throughout the
world. We show that (i) there is significant variance in key performance metrics
both within and across carriers; (ii) this variance is at best only partially explained
by regional and time-of-day patterns; (iii) the stability of network performance
varies substantially among carriers. Further, we use the dataset to diagnose the
causes behind observed performance problems and identify additional measure-
ments that will improve our ability to reason about mobile network behavior.

1 Introduction
Cellular networks are the fastest growing, most popular and least understood Internet
systems. A particularly difficult challenge in this environment is capturing a view of
network performance that is representative of conditions at end user devices. A num-
ber of factors frustrate our ability to capture this view. For instance, carriers enforce
different policies depending on the traffic types or geographic/social characteristics of
different locations such as population [1, 2], causing user perceived performance to dif-
fer from advertised performance for access technologies. Other environmental factors
have a significant impact on performance, including device model [3], mobility [4],
network load [2], packet size [5, 6] and MAC-layer scheduling [4].
To account for various factors impacting Internet performance in mobile networks,
we need pervasive network monitoring that samples a variety of devices across carriers,
access technologies, locations and over time. This work takes a first look at such a view
using data collected from controlled measurement experiments in 144 carriers during
17 months, comprising 11 cellular network technologies. We use this data to identify
the patterns, trends, anomalies, and evolution of cellular networks’ performance.
This study demonstrates that characterizing and understanding the performance in
today’s cellular networks is far from trivial. We find that all carriers exhibit significant

M. Faloutsos and A. Kuzmanovic (Eds.): PAM 2014, LNCS 8362, pp. 12–22, 2014.

c Springer International Publishing Switzerland 2014
 Mobile Network Performance from User Devices 13

variance in end-to-end performance in terms of latency and throughput. To explain this

variance, we investigate geographic and temporal properties of network performance.
While we find that these properties account for some differences in performance, impor-
tantly we observe that performance is inherently unstable, with some carriers providing
relatively more or less predictable performance. Last, we identify alternative sources
of variance in performance that include routing and signal strength. An important open
question is how to design a measurement platform that allows us to understand reasons
behind most observed performance differences.
This paper differs from previous related work in that our study is longitudinal, con-
tinuous, pervasive and gathered from mobile devices using controlled experiments. In
contrast, some related work [7–9] passively collected network traffic from cellular net-
work infrastructure, using one month of data or less. These studies tend to be limited to
a single carrier, hampering our ability to conduct meaningful comparisons across carri-
ers. Other work collected network performance data at mobile devices [10, 1, 11], but
did not use controlled experiments to capture a continuous view of performance.
Roadmap. We describe our methodology and dataset in §2, then present our findings
regarding network performance across different network technologies, carriers, loca-
tions, and times in §3.1, §3.2, and §3.3 respectively. Then we study the root causes for
performance degradation in §3.4. We discuss related work in §4 and conclude in §5.

2 Methodology and Dataset

This paper studies cellular network performance using a broad longitudinal view of net-
work behavior impacting user-perceived performance. To this end, we consider HTTP
GET throughput, round trip time latency from ping, and DNS lookup time as end-to-end
performance metrics. In addition to gathering raw performance data, we annotate our
measurements with path information gathered from traceroute, the identify of the de-
vice’s carrier, its cellular network technology, signal strength, location and timestamp.
We focus on performance from mobile devices to Google, a large, popular content
provider. We argue that Google is an ideal target for network measurements because it is
highly available and well provisioned, making it easier to isolate network performance
to cell networks vs. Google’s network. Focusing on these measurements, we identify the
performance impact of carrier, network technology, location and time. To reason about
the root cause behind performance changes, we use path information, DNS mappings
and signal strength readings.
Our data is collected by two Android apps using a nearly identical codebase,
Speedometer and Mobiperf.1 Speedometer is an internal Android app developed by
Google and deployed on hundreds of volunteer devices, mainly owned by Google em-
ployees. As such, the bulk of our dataset2 is biased toward locations where Google
employees live and work. Speedometer collected the following measurements from
2011-10 to 2013-2 (17 months): 6.6M ping RTTs to www.google.com (each sam-
ple consists of 10 consecutive probes), 1.7M HTTP GETs to measure TCP throughput
1
http://www.mobiperf.com/
2
This dataset is publicly available at
https://storage.cloud.google.com/speedometer
Random documents with unrelated
content Scribd suggests to you:
 copies of the works possessed in a physical medium and
discontinue all use of and all access to other copies of Project
Gutenberg™ works.

• You provide, in accordance with paragraph 1.F.3, a full refund of

any money paid for a work or a replacement copy, if a defect in
the electronic work is discovered and reported to you within 90
days of receipt of the work.

• You comply with all other terms of this agreement for free
distribution of Project Gutenberg™ works.

1.E.9. If you wish to charge a fee or distribute a Project Gutenberg™

electronic work or group of works on different terms than are set
forth in this agreement, you must obtain permission in writing from
the Project Gutenberg Literary Archive Foundation, the manager of
the Project Gutenberg™ trademark. Contact the Foundation as set
forth in Section 3 below.

1.F.

1.F.1. Project Gutenberg volunteers and employees expend

considerable effort to identify, do copyright research on, transcribe
and proofread works not protected by U.S. copyright law in creating
the Project Gutenberg™ collection. Despite these efforts, Project
Gutenberg™ electronic works, and the medium on which they may
be stored, may contain “Defects,” such as, but not limited to,
incomplete, inaccurate or corrupt data, transcription errors, a
copyright or other intellectual property infringement, a defective or
damaged disk or other medium, a computer virus, or computer
codes that damage or cannot be read by your equipment.

1.F.2. LIMITED WARRANTY, DISCLAIMER OF DAMAGES - Except

for the “Right of Replacement or Refund” described in paragraph
1.F.3, the Project Gutenberg Literary Archive Foundation, the owner
of the Project Gutenberg™ trademark, and any other party
distributing a Project Gutenberg™ electronic work under this
agreement, disclaim all liability to you for damages, costs and
expenses, including legal fees. YOU AGREE THAT YOU HAVE NO
REMEDIES FOR NEGLIGENCE, STRICT LIABILITY, BREACH OF
WARRANTY OR BREACH OF CONTRACT EXCEPT THOSE
PROVIDED IN PARAGRAPH 1.F.3. YOU AGREE THAT THE
FOUNDATION, THE TRADEMARK OWNER, AND ANY
DISTRIBUTOR UNDER THIS AGREEMENT WILL NOT BE LIABLE
TO YOU FOR ACTUAL, DIRECT, INDIRECT, CONSEQUENTIAL,
PUNITIVE OR INCIDENTAL DAMAGES EVEN IF YOU GIVE
NOTICE OF THE POSSIBILITY OF SUCH DAMAGE.

1.F.3. LIMITED RIGHT OF REPLACEMENT OR REFUND - If you

discover a defect in this electronic work within 90 days of receiving it,
you can receive a refund of the money (if any) you paid for it by
sending a written explanation to the person you received the work
from. If you received the work on a physical medium, you must
return the medium with your written explanation. The person or entity
that provided you with the defective work may elect to provide a
replacement copy in lieu of a refund. If you received the work
electronically, the person or entity providing it to you may choose to
give you a second opportunity to receive the work electronically in
lieu of a refund. If the second copy is also defective, you may
demand a refund in writing without further opportunities to fix the
problem.

1.F.4. Except for the limited right of replacement or refund set forth in
paragraph 1.F.3, this work is provided to you ‘AS-IS’, WITH NO
OTHER WARRANTIES OF ANY KIND, EXPRESS OR IMPLIED,
INCLUDING BUT NOT LIMITED TO WARRANTIES OF
MERCHANTABILITY OR FITNESS FOR ANY PURPOSE.

1.F.5. Some states do not allow disclaimers of certain implied

warranties or the exclusion or limitation of certain types of damages.
If any disclaimer or limitation set forth in this agreement violates the
law of the state applicable to this agreement, the agreement shall be
interpreted to make the maximum disclaimer or limitation permitted
by the applicable state law. The invalidity or unenforceability of any
provision of this agreement shall not void the remaining provisions.
1.F.6. INDEMNITY - You agree to indemnify and hold the
Foundation, the trademark owner, any agent or employee of the
Foundation, anyone providing copies of Project Gutenberg™
electronic works in accordance with this agreement, and any
volunteers associated with the production, promotion and distribution
of Project Gutenberg™ electronic works, harmless from all liability,
costs and expenses, including legal fees, that arise directly or
indirectly from any of the following which you do or cause to occur:
(a) distribution of this or any Project Gutenberg™ work, (b)
alteration, modification, or additions or deletions to any Project
Gutenberg™ work, and (c) any Defect you cause.

Section 2. Information about the Mission of

Project Gutenberg™
Project Gutenberg™ is synonymous with the free distribution of
electronic works in formats readable by the widest variety of
computers including obsolete, old, middle-aged and new computers.
It exists because of the efforts of hundreds of volunteers and
donations from people in all walks of life.

Volunteers and financial support to provide volunteers with the

assistance they need are critical to reaching Project Gutenberg™’s
goals and ensuring that the Project Gutenberg™ collection will
remain freely available for generations to come. In 2001, the Project
Gutenberg Literary Archive Foundation was created to provide a
secure and permanent future for Project Gutenberg™ and future
generations. To learn more about the Project Gutenberg Literary
Archive Foundation and how your efforts and donations can help,
see Sections 3 and 4 and the Foundation information page at
www.gutenberg.org.

Section 3. Information about the Project

Gutenberg Literary Archive Foundation
The Project Gutenberg Literary Archive Foundation is a non-profit
501(c)(3) educational corporation organized under the laws of the
state of Mississippi and granted tax exempt status by the Internal
Revenue Service. The Foundation’s EIN or federal tax identification
number is 64-6221541. Contributions to the Project Gutenberg
Literary Archive Foundation are tax deductible to the full extent
permitted by U.S. federal laws and your state’s laws.

The Foundation’s business office is located at 809 North 1500 West,

Salt Lake City, UT 84116, (801) 596-1887. Email contact links and up
to date contact information can be found at the Foundation’s website
and official page at www.gutenberg.org/contact

Section 4. Information about Donations to

the Project Gutenberg Literary Archive
Foundation
Project Gutenberg™ depends upon and cannot survive without
widespread public support and donations to carry out its mission of
increasing the number of public domain and licensed works that can
be freely distributed in machine-readable form accessible by the
widest array of equipment including outdated equipment. Many small
donations ($1 to $5,000) are particularly important to maintaining tax
exempt status with the IRS.

The Foundation is committed to complying with the laws regulating

charities and charitable donations in all 50 states of the United
States. Compliance requirements are not uniform and it takes a
considerable effort, much paperwork and many fees to meet and
keep up with these requirements. We do not solicit donations in
locations where we have not received written confirmation of
compliance. To SEND DONATIONS or determine the status of
compliance for any particular state visit www.gutenberg.org/donate.

While we cannot and do not solicit contributions from states where

we have not met the solicitation requirements, we know of no
prohibition against accepting unsolicited donations from donors in
such states who approach us with offers to donate.

International donations are gratefully accepted, but we cannot make

any statements concerning tax treatment of donations received from
outside the United States. U.S. laws alone swamp our small staff.

Please check the Project Gutenberg web pages for current donation
methods and addresses. Donations are accepted in a number of
other ways including checks, online payments and credit card
donations. To donate, please visit: www.gutenberg.org/donate.

Section 5. General Information About Project

Gutenberg™ electronic works
Professor Michael S. Hart was the originator of the Project
Gutenberg™ concept of a library of electronic works that could be
freely shared with anyone. For forty years, he produced and
distributed Project Gutenberg™ eBooks with only a loose network of
volunteer support.

Project Gutenberg™ eBooks are often created from several printed

editions, all of which are confirmed as not protected by copyright in
the U.S. unless a copyright notice is included. Thus, we do not
necessarily keep eBooks in compliance with any particular paper
edition.

Most people start at our website which has the main PG search
facility: www.gutenberg.org.

This website includes information about Project Gutenberg™,

including how to make donations to the Project Gutenberg Literary
Archive Foundation, how to help produce our new eBooks, and how
to subscribe to our email newsletter to hear about new eBooks.