Presentation on theme: "Characterizing Residential Broadband Networks Marcel Dischinger †, Andreas Haeberlen †‡, Krishna P. Gummadi †, Stefan Saroiu* † MPI-SWS, ‡ Rice University,"— Presentation transcript:
Characterizing Residential Broadband Networks Marcel Dischinger †, Andreas Haeberlen †‡, Krishna P. Gummadi †, Stefan Saroiu* † MPI-SWS, ‡ Rice University, * University of Toronto To Be Presented By Muhammad Atif Qureshi, 20093639
2 Marcel Dischinger || IMC 2007 Outline Background and Problem Statement Motivation – Why Study Residential Broadband Networks? Why Do We Know So Little? Finding Broadband Hosts to Measure The Experiment Experimental Setup Experiment Setup and its Validation Experimental Measurements
3 Marcel Dischinger || IMC 2007 Background and Problem Statement – The Last Mile Problem (1/2) LAN, MAN, WAN – how to connect private users at home to such networks? Problem of the Last Mile – somehow connect private homes to the public Internet without laying many new cables. How? By using existing lines – reuse them for data traffic.
4 Marcel Dischinger || IMC 2007 Background and Problem Statement (2/2) Convergence of telephony (voice), Internet (data), cable (video) – gives birth to broadband networks. Broadband networks give best of both worlds. Commonly known as “home networking broadband connection technologies.”
5 Marcel Dischinger || IMC 2007 Motivation - Why study residential broadband networks? Used by millions of users to connect to the Internet Rapidly growing user base Used for many different workloads: Music / movie downloads, VoIP, online games Yet, researchers know little about the characteristics of deployed cable and DSL networks Such as provisioned bandwidths, queueing delays, or loss rates Will have great research implications for designers of future protocols and systems.
6 Marcel Dischinger || IMC 2007 Why do we know so little? Commercial ISPs have no incentives to reveal information about their network deployments Researchers lack access to broadband networks Testbeds composed of academic nodes PlanetLab only has two DSL nodes Prior studies were limited in scale Largest study so far had 47 broadband nodes [PAM’04] Prior studies depended on access to the broadband hosts Challenge: Can we measure hosts without access to them?
7 Marcel Dischinger || IMC 2007 The Experiment - Finding broadband hosts to measure (1/2) Identified IP addresses of broadband hosts using reverse-DNS lookups E.g., BellSouth’s DNS names follow the scheme adsl-*.bellsouth.net Sent TCP ACK and ICMP PING probes to the broadband IPs 1000s of hosts from 100s of DSL/cable ISPs responded
8 Marcel Dischinger || IMC 2007 The Experiment - We focused on 11 major ISPs from North America and Europe (2/2) DSL ISPAmeritechBellSouthPacBellQwestSWBell BT Broadban d RegionS+SW USASE USAS+SW USAW USA S+SW USA UK Hosts Measured 11315515897397173 ISPCharterComcastRoad RunnerRogersChello RegionUSA Canada Netherland s Hosts Measure d 114118301148120 Cable
9 Marcel Dischinger || IMC 2007 Experimental Setup - How do we measure the broadband hosts? We measured from well-connected hosts in University networks TCP ACK / ICMP PING probes sent at 10Mbps for a short duration Probes saturate the bottleneck, which is often the broadband link TCP ACK probes saturate just downstream direction ICMP PING probes saturate both directions We analyzed probe responses to infer various characteristics Internet Broadband link Broadband host Measurement hosts Last-hop router
10 Marcel Dischinger || IMC 2007 Experiment Assumption and its Validation - Are broadband links the bottleneck? Broadband host Last-hop router Broadband links are the bandwidth bottlenecks along the measured path More validation results in the paper
11 Marcel Dischinger || IMC 2007 Rest of the talk Allocated link bandwidths Packet latencies Packet loss
12 Marcel Dischinger || IMC 2007 Outline Allocated link bandwidths Do broadband providers allocate advertised link bandwidths? How do the downstream and upstream bandwidths compare? Are broadband bandwidths stable over the short-term? Are broadband bandwidths stable over diurnal time-scales? Is there evidence for traffic shaping? Packet latencies Packet loss
13 Marcel Dischinger || IMC 2007 Do ISPs allocate advertised link bandwidths? DSL ISPs allocate advertised bandwidths Its dedicated link Some Cable ISPs do not offer discrete bandwidths Its shared link Used asymmetric large-TCP flood PacBell BellSouth Rogers Road Runner
14 Marcel Dischinger || IMC 2007 What is the ratio of downstream to upstream bandwidths? Upstream bandwidths are significantly lower than downstream Broadband networks are provisioned for client-server workloads Used symmetric large ICMP flood for upstreams PacBell Comcast Road Runner Ameritech
15 Marcel Dischinger || IMC 2007 Are link bandwidths stable over the short-term? DSL bandwidths are relatively stable, while cable are not Hard for protocols like TCP to adapt to highly variable cable BWs Unstable (Rogers cable host) Stable (PacBell DSL host)
16 Marcel Dischinger || IMC 2007 Outline Allocated link bandwidths Packet latencies How large are broadband queueing delays? Queues should be proportional to the end-to-end RTT Recent research recommends even shorter queues [SIGCOMM’04] How do cable’s time-slotted policies affect transmission delays? Do broadband links have large propagation delays? Packet loss
17 Marcel Dischinger || IMC 2007 How large are downstream queueing delays? Downstream queues are significantly larger than avg. path RTT Used asymmetric large TCP flood (RTT max - RTT min) PacBell BellSouth Comcast Road Runner
18 Marcel Dischinger || IMC 2007 How large are upstream queueing delays? Upstream queues are extremely large Packets can experience latencies in the order of seconds Used symmetric large ICMP flood – estimated downstream value Greater Upstream Queue length than downstream’s BellSouthPacBell BellSouth Comcast Road Runner
19 Marcel Dischinger || IMC 2007 Why are large queues worrisome? Large queues avoid losses at the cost of latency Good for web workloads But, bad for popular emerging workloads Interactive traffic like VoIP and online games Multimedia downloads like music and movies Low latency vs. maximum bandwidth TCP does not fully drain large queues after a loss event
20 Marcel Dischinger || IMC 2007 Outline Allocated link bandwidths Packet latencies Packet loss Do ISPs deploy active queue management (AQM)? Tail-drop queue Active queue management techniques, such as Random Early Detect (RED) Do broadband links see high packet loss?
21 Marcel Dischinger || IMC 2007 Do ISPs deploy active queue management? 25% of DSL hosts have AQM deployed in the upstream Used small TCP flood Active queue management (probably RED) (SWBell) Tail-drop (PacBell) Threshold
22 Marcel Dischinger || IMC 2007 Conclusion We presented the first large-scale study of broadband networks Measured their bandwidth, latency, and loss characteristics Broadband networks are very different from academic networks Cable networks have unstable bandwidths Large queues can cause latencies in the order of seconds Broadband links have low loss rates, show deployment of AQM Our findings have important implications for network operators and systems designers