Characterizing Residential Broadband Networks

Characterizing Residential Broadband Networks Marcel Dischinger Andreas Haeberlen MPI for Software Systems MPI for Software Systems, Rice University [email protected] [email protected] Krishna P. Gummadi Stefan Saroiu MPI for Software Systems University of Toronto [email protected] [email protected] A large and rapidly growing proportion of users connect to 1. INTRODUCTION the Internet via residential broadband networks such as Dig- Residential broadband networks such as Digital Subscriber ital Subscriber Lines (DSL) and cable. Residential networks Lines (DSL) and cable are increasingly being used to access are often the bottleneck in the last mile of today’s Internet. the Internet. More than 158 million people use these net- Their characteristics critically affect Internet applications, works worldwide [39], and this number is expected to rise including voice-over-IP, online games, and peer-to-peer con- to 477 million by 2011 [51]. In the United States alone, tent sharing/delivery systems. However, to date, few studies more than half of all Internet users connect via residential have investigated commercial broadband deployments, and broadband networks [38]. In addition, many governments rigorous measurement data that characterize these networks are adopting policies to promote ubiquitous broadband ac- at scale are lacking. cess [18,48]. In this paper, we present the first large-scale measurement Residential broadband networks provide the critical ‘last study of major cable and DSL providers in North America mile’ access to the Internet infrastructure. It is widely and Europe. We describe and evaluate the measurement thought that the bottlenecks in the performance of the In- tools we developed for this purpose. Our study character- ternet lie in its access networks [1]. So the reliability and the izes several properties of broadband networks, including link performance of Internet applications – including voice-over- capacities, packet round-trip times and jitter, packet loss IP (VoIP), video on demand, online games, and peer-to-peer rates, queue lengths, and queue drop policies. Our analysis content delivery systems – depend crucially on the charac- reveals important ways in which residential networks differ teristics of broadband access networks. from how the Internet is conventionally thought to operate. Despite the widespread deployment of residential broad- We also discuss the implications of our findings for many band networks and their importance to emerging applica- emerging protocols and systems, including delay-based con- tions, they remain relatively unexplored by the academic gestion control (e.g., PCP) and network coordinate systems community. Although many measurement studies have fo- (e.g., Vivaldi). cused on the Internet’s core [6,26,40] and academic/research edge networks [5,35], rigorous measurement data that char- Categories and Subject Descriptors acterize residential network deployments at scale are lacking. C.2.2 [Computer Systems Organization]: Computer- In the absence of systematic studies, knowledge about res- Communication Networks—Network Operations; C.2.5 idential broadband networks is based on anecdotal evidence, [Computer Systems Organization]: Computer- hearsay, and marketing buzzwords. Although broadband Communication Networks—Local and Wide-Area Networks; networks are known to have very different characteristics C.4 [Computer Systems Organization]: Performance of from academic networks [5, 43], there have been no large- Systems scale studies quantifying these differences. As a result, researchers today are left to second-guess how well protocols or General Terms systems evaluated in academic networks would work in the Measurement, Performance, Experimentation commercial Internet, where broadband networks are widely deployed. Keywords One reason for the lack of large-scale measurement stud- Broadband access networks, DSL, cable, network measure- ies on residential networks is that researchers have limited ment access to broadband environments. Most academic insti- tutions and research laboratories do not access the Inter- net over broadband. Even state-of-the-art research network testbeds, such as PlanetLab [41] and RON [2], have only a Permission to make digital or hard copies of all or part of this work for handful of broadband nodes. We overcame this problem by personal or classroom use is granted without fee provided that copies are developing tools that can measure broadband networks re- not made or distributed for profit or commercial advantage and that copies motely and without cooperation from end hosts connected bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific to the broadband links. permission and/or a fee. In this paper, we present the first large-scale measurement IMC’07, October 24-26, 2007, San Diego, California, USA. study examining 1,894 broadband hosts from 11 major com- Copyright 2007 ACM 978-1-59593-908-1/07/0010 ...$5.00. InternetInternet CMTS Internet Modem CM CM Phone DSLAM Splitter network (a) (b) Figure 1: A typical setup of (a) cable and (b) DSL access networks mercial cable and DSL providers in North America and Eu- 2.1 Cable networks rope. To conduct this study, we developed tools that enable Cable networks use the cable television infrastructure to con- us to measure a large number of remote broadband links. nect home users to the Internet. In these networks, a master We performed a detailed characterization of an extensive set headend connects to several regional headends using fiber- of properties of broadband links. Our analysis was driven optic cables. Each regional headend serves a set of customers by three questions: (up to 2,000 homes). A single coaxial cable, carrying both 1. What are the typical bandwidth, latency, and loss char- television and data signals, links these customers to their acteristics of residential broadband links? headend. DOCSIS [10] is the most common specification defining 2. How do the characteristics of broadband networks dif- the interface requirements of cable modems. In DOCSIS, fer from those of academic or corporate networks? each cable modem (CM) exchanges data with a cable modem 3. What are the implications of broadband-network prop- termination system (CMTS) located in a regional headend. erties for future protocol and system designers? In the downstream direction, the CMTS broadcasts data to all cable modems that are connected to it. The cable Our study reveals important ways in which cable and DSL modems filter all received data and forward only the bytes networks differ from the conventional wisdom about the In- destined for their customer’s host. In the upstream direction, ternet, accumulated from prior studies of academic networks. the access channel is time-slotted – a cable modem must For example, many cable links show high variation in link first reserve a time slot and wait until the CMTS grants the bandwidths over short timescales. Packet transmissions over reservation. When the time slot has been granted, the cable cable suffer high jitter as a result of cable’s time-slotted ac- modem can transmit data upstream. Figure 1(a) illustrates cess policy. DSL links show large last-hop delays and con- a typical setup of a cable access network. siderable deployment of active queue management policies There are several important differences between cable and such as random early detection (RED). Both cable and DSL other access networks. First, cable links typically have asym- ISPs use traffic shaping and deploy massive queues that can metric bandwidths: their downstream bandwidth is much delay packets for several hundred milliseconds. higher than their upstream bandwidth. Second, customers Our findings have important implications for emerging cannot use the full raw capacity of their cable links. Instead, protocols and systems. For instance, the high packet jit- cable networks use traffic shaping to restrict users from con- ter in cable links can affect transport protocols that rely suming more bandwidth than their contract stipulates. Al- on round-trip time (RTT) measurements to detect conges- though cable networks currently allow raw data rates of up tion, such as TCP Vegas [9] and PCP [3]. Further, the large to 40 Mbps, the contracts of individual customers specify queue sizes found in cable and DSL ISPs can be detrimental much lower rates, between 128 Kbps and 10 Mbps. Further, to real-time applications such as VoIP when they are used some ISPs over-subscribe their cable access networks. In this concurrently with bandwidth-intensive applications such as case, the level of service experienced by customers can vary BitTorrent [8]. depending on the amount of competing network traffic. The rest of the paper is organized as follows: Section 2 Finally, cable modems can concatenate multiple upstream provides an overview of residential cable and DSL networks. packets into a single transmission, which results in short Section 3 describes our measurement methodology, including bursts at high data rates. Thus, the upstream latencies can the tools we built for gathering data over remote broadband fluctuate heavily, depending on the allocation policy, and the network links. Section 4 then presents an in-depth analysis of amount of signaling and concatenation used by the CMTS. our data set, characterizing the bandwidth, latency and loss properties of broadband networks. In Section 5, we discuss the implications of our findings

Load more