Measurement and Modeling of Packet Loss in the Internet Maya Yajnik.

Slides:



Advertisements
Similar presentations
Internet Measurement Conference 2003 Source-Level IP Packet Bursts: Causes and Effects Hao Jiang Constantinos Dovrolis (hjiang,
Advertisements

Research Directions Mark Crovella Boston University Computer Science.
Bayesian Piggyback Control for Improving Real-Time Communication Quality Wei-Cheng Xiao 1 and Kuan-Ta Chen Institute of Information Science, Academia Sinica.
Modeling, Simulation and Measurements of Queuing Delay under Long-tail Internet Traffic Michele Garetto Politecnico di Torino, Italy Don Towsley University.
Playback-buffer Equalization For Streaming Media Using Stateless Transport Prioritization By Wai-tian Tan, Weidong Cui and John G. Apostolopoulos Presented.
Doc.: IEEE /0604r1 Submission May 2014 Slide 1 Modeling and Evaluating Variable Bit rate Video Steaming for ax Date: Authors:
Measurements of Congestion Responsiveness of Windows Streaming Media (WSM) Presented By:- Ashish Gupta.
1 Estimating Shared Congestion Among Internet Paths Weidong Cui, Sridhar Machiraju Randy H. Katz, Ion Stoica Electrical Engineering and Computer Science.
Sampling and Flow Measurement Eric Purpus 5/18/04.
End-to-End Routing Behavior in the Internet Vern Paxson Presented by Zhichun Li.
CSE 561 – Multicast Applications David Wetherall Spring 2000.
ACM Multimedia October 4, 2001 Real-time Voice Communication over the Internet Using Packet Path Diversity Yi Liang, Eckehard Steinbach, and Bernd Girod.
1 Estimating Shared Congestion Among Internet Paths Weidong Cui, Sridhar Machiraju Randy H. Katz, Ion Stoica Electrical Engineering and Computer Science.
 Don Towsley 2000 Network Tomography for the Internet: Open Problems D. Towsley U. Massachusetts.
Server-based Inference of Internet Performance V. N. Padmanabhan, L. Qiu, and H. Wang.
Modeling TCP Throughput Jeng Lung WebTP Meeting 11/1/99.
On the Constancy of Internet Path Properties Yin Zhang, Nick Duffield AT&T Labs Vern Paxson, Scott Shenker ACIRI Internet Measurement Workshop 2001 Presented.
1 Network Tomography Venkat Padmanabhan Lili Qiu MSR Tab Meeting 22 Oct 2001.
A simulation-based comparative evaluation of transport protocols for SIP Authors: M.Lulling*, J.Vaughan Department of Computer science, University college.
Network Traffic Measurement and Modeling CSCI 780, Fall 2005.
Monitoring Persistently Congested Internet Links Leiwen (Karl) Deng Aleksandar Kuzmanovic Northwestern University
Available bandwidth measurement as simple as running wget D. Antoniades, M. Athanatos, A. Papadogiannakis, P. Markatos Institute of Computer Science (ICS),
A Nonstationary Poisson View of Internet Traffic T. Karagiannis, M. Molle, M. Faloutsos University of California, Riverside A. Broido University of California,
TAPAS: A Research Paradigm for the Modeling, Prediction and Analysis of Non-stationary Network Behavior Almudena Konrad PhD Candidate at UC Berkeley
Streaming Media. Unicast Redundant traffic Multicast One to many.
1 End-to-End Detection of Shared Bottlenecks Sridhar Machiraju and Weidong Cui Sahara Winter Retreat 2003.
Network Tomography through End- End Multicast Measurements D. Towsley U. Massachusetts collaborators: R. Caceres, N. Duffield, F. Lo Presti (AT&T) T. Bu,
1 Emulating AQM from End Hosts Presenters: Syed Zaidi Ivor Rodrigues.
CSc 461/561 CSc 461/561 Multimedia Systems 0. Introduction.
Medium Start in TCP-Friendly Rate Control Protocol CS 217 Class Project Spring 04 Peter Leong & Michael Welch.
Detecting Shared Congestion of Flows Via End- to-end Measurement Dan Rubenstein Jim Kurose Don Towsley Computer Networks Research Group.
Receiver-driven Layered Multicast Paper by- Steven McCanne, Van Jacobson and Martin Vetterli – ACM SIGCOMM 1996 Presented By – Manoj Sivakumar.
A Hybrid Systems Modeling Framework for Fast and Accurate Simulation of Data Communication Networks João P. Hespanha University of Calif. Santa Barbara.
Slides originally from Professor Williamson at U Calgary1-1 Introduction Part II  Network Core  Delay & Loss in Packet-switched Networks  Structure.
Alok Shriram and Jasleen Kaur Presented by Moonyoung Chung Empirical Evaluation of Techniques for Measuring Available Bandwidth.
Parameter Estimation and Performance Analysis of Several Network Applications Sara Alouf Ph.D. defense - November 8, 2002 Advisor: Philippe Nain.
CS380y Junior Thesis1 The Performance of TCP/IP over Bluetooth Chris Snow Supervisors: Serguei Primak, Electrical Engineering Hanan Lutfiyya, Computer.
Computer Networks: Multimedia Applications Ivan Marsic Rutgers University Chapter 3 – Multimedia & Real-time Applications.
TFRC: TCP Friendly Rate Control using TCP Equation Based Congestion Model CS 218 W 2003 Oct 29, 2003.
Sharing Information across Congestion Windows CSE222A Project Presentation March 15, 2005 Apurva Sharma.
Fluid-based Analysis of a Network of AQM Routers Supporting TCP Flows with an Application to RED Vishal Misra Wei-Bo Gong Don Towsley University of Massachusetts,
1 Passive Network Tomography Using Bayesian Inference Lili Qiu Joint work with Venkata N. Padmanabhan and Helen J. Wang Microsoft Research Internet Measurement.
Hung X. Nguyen and Matthew Roughan The University of Adelaide, Australia SAIL: Statistically Accurate Internet Loss Measurements.
03/11/2015 Michael Chai; Behrouz Forouzan Staffordshire University School of Computing Streaming 1.
Internet Performance Measurements and Measurement Techniques Jim Kurose Department of Computer Science University of Massachusetts/Amherst
Detection of Routing Loops and Analysis of Its Causes Sue Moon Dept. of Computer Science KAIST Joint work with Urs Hengartner, Ashwin Sridharan, Richard.
Multiplicative Wavelet Traffic Model and pathChirp: Efficient Available Bandwidth Estimation Vinay Ribeiro.
WB-RTO: A Window-Based Retransmission Timeout Ioannis Psaras Demokritos University of Thrace, Xanthi, Greece.
Computer Science and Engineering Computer System Security CSE 5339/7339 Session 25 November 16, 2004.
정하경 MMLAB Fundamentals of Internet Measurement: a Tutorial Nevil Brownlee, Chris Lossley, “Fundamentals of Internet Measurement: a Tutorial,” CMG journal.
End-to-End Routing Behavior in the Internet Vern Paxson Presented by Sankalp Kohli and Patrick Wong.
Development of a QoE Model Himadeepa Karlapudi 03/07/03.
Internet research Needs Better Models Sally Floyd, Eddie Kohler ISCI Center for Internet Research, Berkeley, California Presented by Max Podlesny.
An Extensible RTCP Control Framework for Large Multimedia Distributions Paper by: Julian Chesterfield Eve M. Schooler Presented by: Phillip H. Jones.
1 Advanced Transport Protocol Design Nguyen Multimedia Communications Laboratory March 23, 2005.
Measurement and Modelling of the Temporal Dependence in Packet loss Maya Yajnik, Sue Moon, Jim Kurose, Don Towsley Department of Computer Science University.
-Mayukh, clemson university1 Project Overview Study of Tfrc Verification, Analysis and Development Verification : Experiments. Analysis : Check for short.
1 Internet Traffic Measurement and Modeling Carey Williamson Department of Computer Science University of Calgary.
1 Network Tomography Using Passive End-to-End Measurements Lili Qiu Joint work with Venkata N. Padmanabhan and Helen J. Wang.
Access Link Capacity Monitoring with TFRC Probe Ling-Jyh Chen, Tony Sun, Dan Xu, M. Y. Sanadidi, Mario Gerla Computer Science Department, University of.
A special acknowledge goes to J.F Kurose and K.W. Ross Some of the slides used in this lecture are adapted from their original slides that accompany the.
2: Transport Layer 11 Transport Layer 1. 2: Transport Layer 12 Part 2: Transport Layer Chapter goals: r understand principles behind transport layer services:
1 Ad-hoc Transport Layer Protocol (ATCP) EECS 4215.
Probabilistic Congestion Control for Non-Adaptable Flows Jörg Widmer, Martin Mauve, Jan Peter Damm (NOSSDAV’02) Presented by Ankur Upadhyaya for CPSC 538A.
Chapter 9: Transport Layer
Monitoring Persistently Congested Internet Links
An IP-based multimedia traffic generator
Empirically Characterizing the Buffer Behaviour of Real Devices
Pong: Diagnosing Spatio-Temporal Internet Congestion Properties
Privacy-Preserving Dynamic Learning of Tor Network Traffic
Presentation transcript:

Measurement and Modeling of Packet Loss in the Internet Maya Yajnik

Overview Context and motivation Contributions of my thesis Loss in the MBone multicast network Temporal correlation of loss Accuracy of loss measurements Summary

Network Protocol Design Providing reliability, congestion control, flow control for –multimedia applications –multicast networking Multimedia traffic in the Internet –streaming multimedia: web-based audio/ video clips –interactive multimedia: Internet telephony, audio/video conferencing

Multicast Networking allows group communication application: audio/ video conferencing MBone: multicast backbone overlaid over the Internet –experimental testbed for application design

Why measure and model loss? Understanding underlying network behavior leads to informed design choices Observations and models useful in analysis and simulation of performance of network protocols Useful to –characterize general patterns of network behavior –find where in the network impairments occur –detect anomalous behavior

Contributions of My Thesis Loss in the MBone multicast network: –estimated where loss occurs in the network –modeled spatial correlation in loss –characterized loss bursts Temporal correlation of loss: –estimated correlation timescale of loss –modeled temporal correlation in loss Accuracy of probe loss measurements: –found they capture congestion level –found they do not capture overall loss rate

Measurement of Loss in MBone Sender transmits audio data at regular intervals Data collecting programs at receivers give end-end behavior 17 geographically distributed receivers off-line analysis of data

Internet Topology Backbone Edge

Where does MBone loss occur? Methodology: –link loss inferred from loss at receivers –correlation of received packets provides glimpse inside Results: –observable backbone loss small 0.01% 0.1% 0.002% 0.2% 0.01% 0.4% 0.2% 7% 0.1% 1% 0.4%16% 21% 0.1% 0.01% 0.04% 0.5% 5% California Mass. Sweden Germany Texas Virginia France Maryland Kentucky Cal. Wash.

Simultaneous Loss and Models Models of Spatial Correlation –star topology –full topology –modified star topology

Loss Burst Characterization Question: do losses occur singly or in long bursts? Results: –mostly singly –occasional long periods of 100% loss lasting 10 seconds to 2 minutes

Summary: Multicast Loss Measured loss at 17 geographically distributed sites in the MBone multicast network Inferred link loss from loss at receivers Backbone loss found to be small Modified star found to be a good model Most losses occur singly Occasional long outages

Context and motivation Contributions of my thesis Loss in the multicast network Temporal Correlation of loss Accuracy of loss measurements Summary Overview

Time Correlation in End-end Loss Questions: –what is the time correlation of packet loss? –what is good model for the loss process? Useful for: –design, performance analysis and simulation adaptive mechanisms for multimedia applications (eg. coding techniques) on-line loss estimation in multimedia applications

Temporal Correlation Internet time lag Observations at the receiver 4521 loss

Temporal Correlation Overview Measurement Analysis –stationarity –data representation –temporal correlation modeling –Markov chain models –estimation of order Summary

Measurement Methodology collected point-point, multicast traces of periodically generated probes probes sent at regular intervals of 20ms, 40ms, 80ms, 160ms source: University of Massachusetts Amherst destinations: Atlanta, Los Angeles, Seattle, St. Louis, Stockholm 128 hours of data

Stationarity Divided trace into 2 hour segments Checked for stationarity –look for change in loss average over trace –removed non-stationary sections Result: selected 76 hours of data

Data Representations binary time series –no loss: 0, loss: 1 –eg. { } interleaved sequences of good run lengths, loss run lengths –eg.{ } {3,5} {2,1} good loss good loss {{{{

Correlation Timescale goal: time interval between packets, at and beyond which loss events are independent methodology: –autocorrelation function –95% bounds around zero for sampling error –chi-square test for independence

Correlation Timescale finding: correlation timescale usually 1 second, often < 640ms

Run lengths: Correlation question –are they independent? methodology –autocorrelation functions, –crosscorrelation function findings –160ms traces: independent –20ms,40ms traces: dependent good runs

question –how are they distributed? –geometrically ? good run length distribution loss run length distribution Run lengths: Distributions

Models We propose using –k-th order Markov chain models –prob. of loss/no loss depends k previous events (i.e. the state) –number of states = 2 k Previously used: –Bernoulli loss (order 0): independent loss –2-state model (order 1): prob. of loss/no loss depends on the previous event order 1 model order 2 model

Order of the Markov process relevant history order 3 Markov process correlation timescale = 640ms For an example 160ms trace

Models Question: what is the appropriate order of the Markov process? –the lag beyond which the loss events are “independent” –related to correlation timescale Results: –160ms traces: order 0 (Bernoulli) : 14 hr / 66 hr order 1 (2-state model): 20 hr/ 66 hr order 2-6: 32 hr/ 66 hr –40ms traces: order 10, 14, 22 –20ms traces: order 17, 42

Temporal Correlation Summary collected/ analyzed 128 hours of loss data correlation timescale < 1000ms Markov chain models of k-th order Bernoulli or 2-state models accurate for aproximately half the data

Accuracy of probe loss measurements Stream of packets “probe” the state of the network (congested or not) UDP datagram probes Periodic Probes Poisson Probes

Accuracy of loss measurements Questions: Does probe loss rate reflect congestion level in the network? –Answer: yes –no appreciable difference between periodic and Poisson probes Does probe loss rate reflect the overall packet loss rate of traffic? –Answer: no

Methodology Network simulation –can record network state and performance congestion level probe loss rate –probing intervals 1ms to 100ms overall packet loss rate measure of probe performance normalized difference between probe loss rate and congestion level

Simulation Topology Bottleneck link –1Mbps and 10Mbps –buffer size of 50 packets –focus on forward direction only Traffic –TCP sessions –on-off sources

Simulation Methodology Congestion level –average fraction of time bottleneck queue is full Probe traces –sample state of the queue –binary sequences: eg –0: queue is not full, 1: queue is full –no packets sent

Sampling network state baseline periodic samples baseline Poisson samples select subset of samples

Results Question: does probe loss rate capture the congestion level? Measure: Error in probes’ estimation of congestion level

Results Question: Does probe loss rate capture the overall packet loss rate?

Summary: Accuracy of loss measurements Questions: Does probe loss rate reflect congestion level in the network? –Answer: yes –no appreciable difference between periodic and Poisson probes Does probe loss rate reflect the overall packet loss rate of traffic? –Answer: no

Contributions of My Thesis Loss in the MBone multicast network: –estimated where loss occurs in the network –modeled spatial correlation in loss –characterized loss bursts Temporal correlation of loss: –estimated correlation timescale of loss –modeled temporal correlation in loss Accuracy of probe loss measurements: –found they capture congestion level –found they do not capture overall loss rate