Performance Limitations of ADSL Users A Case Study

Performance Limitations of ADSL Users: A Case Study Matti Siekkinen, University of Oslo Denis Collange, France Télécom R&D Guillaume Urvoy-Keller, Ernst W. Biersack, Institut Eurecom PAM April 6, 2007

Outline q Introduction § Motivation § Techniques for root cause analysis of TCP throughput q Measurement setup q Analysis results q Conclusions 2/22/2021

Introduction q What? § Analyzed 24 h packet trace from France Telecom’s ADSL access network § Studied throughput limitations experienced by clients q Why? § Knowing throughput limitations (=performance) is useful o o ISPs want satisfied clients Need to know what’s going on before things can be improved q How? § Root Cause Analysis of TCP Throughput Analysis and inference of the reasons that prevent a given TCP connection from achieving a higher throughput. o Passive traffic analysis o § Why TCP? o TCP typically over 90% of all traffic 2/22/2021

Background q “On the characteristics and origins of Internet flow rates” by Zhang et al. (SIGCOMM 2002) § Pioneering research work § Congestion is not always the cause for throughput limitations 2/22/2021

Limitation Causes for TCP Throughput q Application § The application does not even attempt to use all network resources § E. g. streaming applications and “bursty” applications (Web browsing) q Transport layer § TCP receiver o o Receiver advertized window limits the rate max amount of outstanding bytes = min(cwnd, rwnd) Flow control Configuration issue • default receiver advertized window is set too low • window scaling is not enabled § TCP protocol o Ramp-up period in slow start and congestion avoidance q Network layer § Congestion at a bottleneck link 2/22/2021

Measurement Setup access network collect network Internet Two pcap probes here q 24 hours of traffic on March 10, 2006 § Passively capture all TCP/IP headers analyze offline q 290 GB of TCP traffic § 64% downstream, 36% upstream q Observed packets from ~3000 clients, analyze only 1335 § Excluded clients did not generate enough traffic for RCA 2/22/2021

Warming up… q Connections § Size distribution highly skewed § Use only 1% of the flows for RCA o Represent > 85% of all traffic q Clients § Heavy-hitters: 15% of clients generate 85 -90% of traffic (up & down) § Low access link utilization 2/22/2021

Results of Limitation Analysis contains most bytes contains some bytes q Few active clients overall q Application limitation dominates q Network limitation by distant bottleneck also experienced 2/22/2021

Application analysis: Application limited traffic other e. Donkey q Quite stable and symmetric volumes § Vast majority of all traffic q e. Donkey and “other” dominate 2/22/2021 P 2 P

Application analysis: Saturated access link q No recognized P 2 P q Asymmetric port 80/8080 downstream § Real Web traffic? 2/22/2021

Impact of Limitation Causes q How far from optimal (access link saturation) are we? q Main observations § Very low downlink utilization for application limited traffic o § Utilization < 20% during 65% of application limited periods of traffic Uplink utilization < 50% during most of application and network limited uploads 2/22/2021

Connecting the evidence… q Most clients’ performance limited by applications q Very low link utilizations for application limited traffic q Most of application limited traffic seems to be P 2 P § Peers often have asymmetric uplink and downlink capacities § P 2 P applications/users enforce upload rate limits Poor aggregate download performance Interne t downloading client Low downlink utilization 2/22/2021 Low uplink capacity +rate limiter uploading clients

Conclusions q Analyzed 24 h packet trace from France Telecom’s ADSL access network § Studied throughput limitations experienced by clients q Majority of clients mostly throughput limited by applications § P 2 P clients throttle upload rate o § Too much? Asymmetric link capacities q Impact and implications § § ISP traffic is mostly application limited traffic Things can change dramatically with o o More intelligent P 2 P clients Caches 2/22/2021

For the future… q Play with time scale q Extended case study on ADSL clients § We saw a day, what about a week? q Could we do things on-line? q Improving RCA techniques § § Short connections Non FIFO traffic (e. g. wireless) 2/22/2021

Thank you for your attention