PAM 2007 siekkine

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: 

Outline Introduction Motivation Techniques for root cause analysis of TCP throughput Measurement setup Analysis results Conclusions

Introduction: 

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

Background: 

Background “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

Limitation Causes for TCP Throughput: 

Limitation Causes for TCP Throughput Application The application does not even attempt to use all network resources E.g. streaming applications and “bursty” applications (Web browsing) Transport layer TCP receiver 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 Ramp-up period in slow start and congestion avoidance Network layer Congestion at a bottleneck link

Measurement Setup: 

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

Warming up…: 

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

Results of Limitation Analysis: 

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

Application analysis:Application limited traffic: 

Application analysis: Application limited traffic Quite stable and symmetric volumes Vast majority of all traffic eDonkey and “other” dominate P2P other eDonkey

Application analysis:Saturated access link: 

Application analysis: Saturated access link No recognized P2P Asymmetric port 80/8080 downstream Real Web traffic?

Impact of Limitation Causes: 

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

Connecting the evidence…: 

Connecting the evidence… Most clients’ performance limited by applications Very low link utilizations for application limited traffic Most of application limited traffic seems to be P2P Peers often have asymmetric uplink and downlink capacities P2P applications/users enforce upload rate limits  Poor aggregate download performance Internet Low downlink utilization Low uplink capacity +rate limiter downloading client uploading clients

Conclusions: 

Conclusions Analyzed 24h packet trace from France Telecom’s ADSL access network Studied throughput limitations experienced by clients Majority of clients mostly throughput limited by applications P2P clients throttle upload rate Too much? Asymmetric link capacities Impact and implications ISP traffic is mostly application limited traffic Things can change dramatically with More intelligent P2P clients Caches

For the future…: 

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

Thank you for your attention: 

Thank you for your attention

Backup slides: 

Backup slides

Our approach (suppress): 

Our approach (suppress) Analyze passive traffic measurements Capture and store all TCP/IP headers, analyze later off-line Observe traffic only at a single measurement point Applicable in diverse situations E.g. at the edge of an ISP’s network Know all about clients’ downloads and uploads Bidirectional packet traces Connection level analysis

Scope (suppress): 

Scope (suppress) Study long lived TCP connections Short connections are another topic Dominated by slow start? Assume FIFO scheduling Necessary for link capacity estimations with packet dispersion techniques Reasonable assumption for most traffic May not hold for cable modem and 802.11 access networks

Warming up…: 

Applications Port based identification Connections Size distribution highly skewed Use only 1% of them for RCA Represent > 85% of all traffic Clients Heavy-hitters: 15% of clients generate 85-90% of traffic (up & down) Low access link utilization Why? Warming up… >5% of traffic each

Client-level root cause analysis: 

Client-level root cause analysis Limitation causes for clients Application Saturated access link Network limitation due to distant bottleneck link TCP configuration Connection-level RCA ALPs network limited BTPs during which utilization > 90% network limited BTPs during which utilization < 90% download (=local problem) BTPs limited by TCP layer Extend the InTraBase framework

Results of Limitation Analysis: 

Results of Limitation Analysis Few active clients overall Application limitation dominates Network limitation by distant bottleneck also experienced similar contains most bytes contains some bytes

Application analysis:Distant bottleneck link: 

Application analysis: Distant bottleneck link Diverse mixture Cause is not necessarily due to client’s behavior

Impact of Limitation Causes: 

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

Impact of Limitation Causes: 

Impact of Limitation Causes Very low downlink utilization for application limited traffic upstream downstream How far from optimal (access link saturation) are we?