Uncovering Cellular Network Characteristics: Performance ... hjx/file/CSE-TR-578-13.pdfUncovering Cellular Network Characteristics: Performance, Infrastructure, and Policies ... network

  • View

  • Download

Embed Size (px)

Text of Uncovering Cellular Network Characteristics: Performance ... hjx/file/CSE-TR-578-13.pdfUncovering...

  • Uncovering Cellular Network Characteristics:Performance, Infrastructure, and Policies

    Junxian Huang Feng Qian Qiang Xu

    Zhiyun Qian Z. Morley Mao Ammar Rayes University of Michigan Cisco

    ABSTRACTMobile Smart Devices (Smartphones and tablets) have be-come increasingly popular especially in IT based servicemanagement companies. According to IDC, more than 70%of executives and sale managers are replacing their PCswith tablets. This is in part due to the agility flexibilityand the availability of diverse network-based and supportapplications. Thus network characteristics directly affectuser-perceived performance and a deep understanding of theproperties of contemporary cellular networks for commonlyused platforms is important for smartphone application andplatform optimization.

    In this work, we carry out the largest study to date of cel-lular networks in terms of users, time duration, location,and networks to understand the performance, infrastruc-ture, and policy characteristics. With the data set collectedfrom around 100K users across the world over 18 months,MobiPerf, a smartphone network measurement tool wedeveloped and publicly deployed, enables us to analyze net-work performance along several new dimensions, previouslynot examined. Our results indicate that with better infras-tructure support, large cities appear to have better perfor-mance than rural areas. Our case study on packet sizes ef-fect on RTT uncovers a surprising pattern for AT&Ts uplinkRTT. We also show that Internet-based CDN service pro-vides very limited latency improvement in todays cellularnetworks. We further examine how local DNS servers areassigned to mobile users. In addition, we scrutinize the car-riers policy towards different types of traffic and success-fully identify some middlebox behavior of todays cellularcarriers.

    1. INTRODUCTIONGiven the wide adoption of smartphone platforms, such

    as iOS and Android, there is a growing number of popu-lar mobile applications designed for these platforms. Formany of these applications, including web browser, email,VoIP, social networks, network access is required. Even forgames that are often run locally, ranking systems and on-line peer matching systems are widely adopted which alsorequires network access, e.g., Game Center for iOS. As aresult, mobile data traffic volume is sky-rocketing. For ex-ample, AT&Ts mobile data volumes surged by a staggering8,000% from 2007 to 2010 [1]. Given the limited networkresources available, it would not be surprising if a carrier

    enforces different policies depending on the traffic types orusers. Hence, it is critical to understand performance, infras-tructure, and policy in cellular networks.

    Our previous study [11] comparing cellular network per-formance among different carriers has already shown indica-tion that network can be the bottleneck accounting for poorapplication performance. In this work, with a data set col-lected for a much longer period (18 months) and a largeruser set (about 10,000 unique users across three major smart-phone platforms), we study the correlation between perfor-mance and several important new dimensions, including net-work types, location, time, etc. to delve deeper into cellu-lar network behavior. In addition, we have conducted localexperiments to understand how packet size affects end-to-end latency. Our study uncovers important characteristicsfor cellular network performance.

    We are motivated by our previous study on cellular net-work infrastructure [19]. In this study, with a comprehensivelatency measurement data set, we perform more in-depthanalysis and quantify the effectiveness of CDN servers forcellular networks. We also provide fine-grained analysis ofthe geographical coverage of each individual IP address oflocal DNS (LDNS) servers and discuss the implications.

    Traffic differentiation has long been studied in the Inter-net [20, 9] given the controversy surrounding the idea ofnetwork neutrality. For mobile networks, most policies incellular networks remain unknown. Our effort shares thesame goal as the WindRider [5] project for monitoring mo-bile network neutrality, but we are the first to report someconclusive results. In this work, we make one of the firstattempts to uncover these policies for cellular carriers thataffect application performance.

    The key contributions and results in this work include:

    We compare the performance across different mobilenetwork technologies (WiFi, 3G, EDGE, GPRS) andwithin 3G technologies (1xRTT, EVDO, UMTS, HS-DPA), providing the largest scale and most comprehen-sive performance comparison to date.

    We analyze the correlation between RTT and packetsize, and find that uplink latency for AT&Ts 3G net-work is a step function of packet size, in contrast topacket size independent RTT behavior for T-Mobile.

    We show little correlation between network latencyand physical distance, and rather limited effectivenessof CDN service in todays cellular networks.


  • We study traffic policy of cellular carriers and success-fully detect some middlebox behavior for T-Mobile.With data collected from global users, we also makeone of the first studies of middlebox behavior acrosslocations and carriers.

    The remainder of this paper is organized as follows: In2, we discuss the methodology for local experiments. Thenin 3, results related with cellular network performance arediscussed, followed by our study on network infrastructurein 4. We discuss policy in cellular networks in 5, beforesummarizing related work in 6 and concluding in 7.

    2. METHODOLOGYIn this study, we use measurement data collected from a

    publicly deployed tool MobiPerf1 as well as from localexperiments. MobiPerf is designed to collect anonymizednetwork measurement information directly from end users,including network type, carrier, GPS, latency to landmarkserver, as well as TCP and DNS performance. The keymethodology for designing MobiPerf has been discussedin our previous work [11, 19]. In this section, we discuss theimprovements and the setup for local experiments.

    We have conducted two major sets of local experiments.These tests are implemented as Android applications run-ning on our Android smartphones locally. For both exper-iments, at every second, the device sends and receives alarge packet (MTU) to ensure the radio interface is at highpower state to occupy the high speed data transmission chan-nel [15].RTT vs. Packet Size We study the correlation betweenRTT and packet size for different carriers using TCP andUDP (for both uplink and downlink) to explore possible ef-fect of packet size on end-to-end delay. The packet size in-creases from 100 bytes (including headers) with an incre-ment of 25 bytes. The response packet has 1 byte payload tofocus on a single direction at a time.Port Scanning We use three Android smartphones withAT&T, T-Mobile, and Verizon 3G service enabled respec-tively. The ports scanned are either popular Internet portsor special ports for mobile platforms, e.g., port 5228 is usedby various Android services including Android Market, andport 5223 is used by Apples push notification service. Foreach port, there is a TCP server and a UDP server run-ning on a local host to simply echo back any message re-ceived. At the client side, an Android app first connects tothe TCP server by sending a short (100 bytes including head-ers) unique message. The client then sends another shortunique message to the UDP server. The client measures thetime to establish a TCP connection, to get the response backin TCP or UDP. Besides UDP port 161 (SNMP), no otherport is blocked by the firewall at the server side. We do notmeasure TCP data transfer time for TCP ports 22, 80 and1MobiPerf is the newer version the measurement tool 3GTest wedeveloped, with various functionality and UI improvements

    443 for simplicity. All ports are scanned sequentially withthe entire scanning process repeated for more than 48 hours.


    We first describe the public deployment of our mobile net-work measurement tool, and then present the performanceanalysis along dimensions including technology type, time,and location. In addition, we study in more depth the corre-lation between RTT and packet size in cellular networks.

    3.1 MobiPerf Deployment and User StatisticsWe publicly deployed the MobiPerf application in Au-

    gust, 2009, distributed via Apples App Store, GooglesAndroid Market and Microsofts Windows Marketplace forMobile. Ever since the initial deployment, we have beencontinuously improving and releasing updates for iOS andAndroid version of our app. Till April, 2011, 99.1K usersfrom across the world have run our app for 439.5K times.The number of users and runs for three different platforms,including iOS, Android, and Windows Mobile, is listed inTable 1. The average number of runs for each Android useris larger than the other two platforms, because for the An-droid version of our app, we give an option to the users toperiodically run the tests. We observe users from 179 coun-tries or regions according to the collected GPS information.Among all 93.3K users, 63.7K (68.27%) have GPS readingsand 52.24% of them are from the U.S., and among these63.7K users, about 1.0K (1.57%) users have run our app inmore than one countries or regions. We also observe morethan 800 carrier names. However, carriers may adopt differ-ent names in different countries, making it difficult to accu-rately estimate the actual number of carriers. Figure 2 showsthe user coverage of MobiPerf, with one dot representingone run of MobiPerf. Given the wide coverage of regions,we believe our data set is fairly representative of the entiresmartphone population, especially for North America withdenser user distribution. In this study, our analysis mostlyfocuses on U.S.