A New Perspective on Internet Quality of Service: Measurement and Predictions Soshant Bali*, Yasong Jin**, Victor S. Frost* and Tyrone Duncan** Information and Telecommunication Technology Center *Electrical Engineering & Computer Science **Department of Mathematics frost@eecs.ku.edu, 785-864-4833 University of Kansas A KTEC Center of Excellence 1 What is the perceived QoS for this end-to-end path? University of Kansas A KTEC Center of Excellence 2 Outline • Develop end-to-end measurement techniques • Develop prediction methodologies for fBM traffic • A Few Words about our Graduate and Research Programs at EECS@KU University of Kansas A KTEC Center of Excellence 3 Premise • Voice networks had a very understandable QoS metric-Blocking • Internet QoS metrics must correlate to end-user experience. • Metrics such as delay and loss may have little direct meaning to the end-user because knowledge of specific coding and/or adaptive techniques is required to translate delay and loss to the user-perceived performance. • Detecting “observable impairments” must be independent of coding, adaptive playout or packet loss concealment techniques employed by the multimedia applications. • Time between impairments and their duration are metrics that are easily understandable by network user. • This research developed methods to detect these impairment events using end-to-end measurements. University of Kansas A KTEC Center of Excellence 4 Network states • Noticeable impairments for Real-time multimedia (RTM) services occur when the endto-end connection is in one or more of the following states: • • • • Burst loss, High random loss, Disconnected, High Delay. • Two other connection states are defined: • Congested, • Route change. University of Kansas A KTEC Center of Excellence 5 Background • End-to-end argument • end nodes: most functions implemented here including application specific functions • core: important forwarding and routing functions are implemented here; not burdened by application specific functions, e.g., reliable delivery • Anomalous events • failures: fiber cuts, power failures etc. • congestion • cause user-perceived impairments • Inferring anomalous events from end-to-end observations • core nodes implement simple functions; do not inform end nodes of anomalous events • need to infer anomalous events from end-to-end observations • Several benefits if anomalous events are accurately inferred University of Kansas A KTEC Center of Excellence 6 Significance • A new QoS metric for RTM applications • ISPs can use impairments metric in service level agreements (SLAs) • Fault diagnosis tools for ISPs • an alternative to traceroute for detecting layer 3 route changes • method for detecting layer 2 failures • Routing for overlay / content delivery networks • Increasing TCP throughput • Confidence interval for minimum RTT estimate (byproduct) University of Kansas A KTEC Center of Excellence 7 Goal • Given a set of active end-to-end network measurements determine the network state and the temporal characteristics of impairment events Network Round Trip Time Packet Loss Rate Traceroute Time-to-live Network State Impairment Events: -Frequency -Duration University of Kansas A KTEC Center of Excellence 8 Goal University of Kansas A KTEC Center of Excellence 9 Route Change • Motivation • Route changes can cause user perceived impairments • Need to divide observations into “homogenous” regions • Layer 3 route changes • • • • TTL Traceroute Not all route changes result in TTL change Not all routers respond to ICMP massages for traceroute • Layer 2 route changes are not visible end-toend University of Kansas A KTEC Center of Excellence 10 Route change state • RTT based route change detection • in figure below, minimum RTT changed but traceroute and TTL field of IP header did not change; layer 2 route change • • • TTL change: not all route changes result in TTL change traceroute change: inefficient, not all routers respond to ICMP massages for traceroute both layer 2 and layer 3 route changes can be detected using RTT based route change detection University of Kansas A KTEC Center of Excellence 11 Route Change Layer 2 Route Change If – the time between changes > ΔT – and the RTT difference across the route change > ΔRTT – and variation in RTT<VRTT – Then the proposed algorithm can detect the change Route Change detected using the discussed procedure (planetlab1.cambridge.intel-research.net and planet1.berkeley.intel-research.net, August 2004) University of Kansas A KTEC Center of Excellence 12 Congested State Observed from M/M/1 Queues The end-to-end flow is in the Congested sate if: Where = Ave waiting time is an indicator of congestion = Packet loss rate University of Kansas A KTEC Center of Excellence 13 Congested State RTTs and a congestion event detected using the discussed procedure planetlab2.ashburn.equinix.planet-lab.org and planetlab1.comet.columbia.edu, 2/04 University of Kansas A KTEC Center of Excellence 14 Delay Impairment State • Given the RTT data, an estimate is made of the minimum playout delay buffer size that is needed to avoid excessive packet losses. • If minimum playout delay > Dplayout then a Estimated one-way delays and minimum playout delay delay impairment has planetlab2.ashburn.equinix.planet-lab.org occurred. and planetlab1.comet.columbia.edu Feb, 2004 University of Kansas A KTEC Center of Excellence 15 Other Networks States • Disconnected state • Period of consecutive packet losses > Ψ sec • Burst loss state • ξsec < Period of consecutive packet losses < Ψ sec • High Random Loss State • Insure enough observed losses, e.g., N, for “valid” loss probability estimate, RoT N > 10 • Observe N losses, if number of packets between the first and Nth loss < NL then network in high lose state University of Kansas A KTEC Center of Excellence 16 Measurement data University of Kansas A KTEC Center of Excellence 17 Congestion Events observed over a period of one week (DC1) University of Kansas A KTEC Center of Excellence 18 Statistics of user-perceived impairments University of Kansas A KTEC Center of Excellence 19 Other observations • Layer 2 route change • 96 events were manually classified as layer 2 route changes • ~71.8% layer 2 route changes were detected by the algorithm • ~4% of the detected events were false positives. • ~8% of all layer 3 route changes were preceded by burst or disconnect loss events. University of Kansas A KTEC Center of Excellence 20 Summary of measurement results • mean time between impairments: from 3.52hrs to 268hrs • mean duration of impairments: from 4.4mins to 92.5mins • on 2 paths congestion for 6-8 hrs during day (weekdays) • • burst loss, high random loss and high delay events were observed when connection was in congested state mean time between burst loss events that occurred during congestion = 14 min, mean duration = 22.64 sec • mean time between layer 3 route changes = 7.23 hrs • • • 18% of all layer 3 route changes 1 sec apart, 15% 2 sec apart, 80% less than 45 mins apart 8% of all layer 3 route changes were preceeded by burst or disconnect loss events mean duration of burst loss events that precede layer 3 route changes = 113.5 sec • mean time between layer 2 route changes = 58.22 hrs • none of the layer 2 route changes were preceded by burst loss events University of Kansas A KTEC Center of Excellence 21 Experimental Conclusions • Developed procedures to detect impairment states for RTM services using end-to-end measurements. • Developed techniques to detect layer two route changes and congestion • The developed techniques consider multiple metrics at the same time to infer the presence of user perceived impairments. Details in “Characterizing User-perceived Impairment Events Using End-to-End Measurements, Soshant Bali, Yasong Jin, V. S. Frost and T. Duncan, International Journal of Communication Systems. University of Kansas A KTEC Center of Excellence 22 Queue Size in Bits Predicting Properties of Congestion Events University of Kansas A KTEC Center of Excellence 23 Queue Size in Bits Predicting Properties of Congestion Events University of Kansas A KTEC Center of Excellence 24 Predicting Properties of Congestion Events • • • • • • • Traffic Model fractional Brownian motion (fBm) Qo(t) = Queue length at t m=Service rate m=average input rate a=variance of the input rate BH(t)=standard fBm with parameter H c=scaled surplus rate University of Kansas A KTEC Center of Excellence 25 Sojourn Time University of Kansas A KTEC Center of Excellence 26 Inter congestion event time University of Kansas A KTEC Center of Excellence 27 Congestion duration University of Kansas A KTEC Center of Excellence 28 Amplitude University of Kansas A KTEC Center of Excellence 29 Conclusions • Developed methods to measure impairments using end-to-end measurements • Developed techniques to predict several properties of congestion events for fBM traffic: • • • • Rate, Duration, Amplitude For details see: “Predicting Properties of Congestion Events for a Queueing System with fBM Traffic”, Yasong Jin, Soshant Bali. Tyrone Duncan, Victor S. Frost, accepted pending revisions for the IEEE Transactions on Networking. University of Kansas A KTEC Center of Excellence 30 A Few Words about our Graduate Program at EECS@KU • 37 faculty • • • • 4 Fellows of the IEEE Ex-Program Managers from DARPA, NSF, NASA 10 new faculty in the past 3 years Currently recruiting one more faculty member • 150 MS students • 75 Ph.D. students • Almost all our Ph.D. students are supported • MS degrees in EE, CoE, CS • Ph.D. degrees in EE, CS • Two major research labs: ITTC and CReSIS • Research volume of over $20 million, with research expenditures of $5.5 million in 2005 • >50% of our graduate students are supported (over 140 in F’05) University of Kansas A KTEC Center of Excellence 31 EECS Research Space University of Kansas A KTEC Center of Excellence 32 What Some of Our Recent Graduates Are Doing Now • Cory Beard (PhD EE 1999) – Associate Professor UMKC • Jennifer Leopold (PhD CS 2000) - Professor of CS at Missouri, Rolla • Amit Kulkarni (PhD CS 2000) - GE Global Research Center • Daniel Cliburn (PhD CS 2001) - Professor of CS at Hanover College • Nathan Goodman (PhD EE 2002) - Professor of ECE at the University of Arizona • Cindy Kong (PhD CS 2004) - Intel Corp. • Wesam Alanqar (PhD EE 2005) - Sprint Corp. • Jungwoo Ryoo (PhD CS 2005) - Professor at Arizona State University • David Janzen (PhD CS 2006) - Professor at Cal Poly, San Louis Obispo University of Kansas A KTEC Center of Excellence 33 Ph.D. Focus Areas • • • • • Communication Systems and Networking Computer Systems Design Interactive Intelligent Systems Bioinformatics Radar Systems and Remote Sensing University of Kansas A KTEC Center of Excellence 34 Communication Systems and Networking • Advancing knowledge of systems interconnected via radio and other technologies • New methodologies to determine the performance and protection of Internetbased systems • Theory and technologies that enable the delivery of reliable information in support of end-user applications independent of the access technology University of Kansas A KTEC Center of Excellence 35 Computer Systems Design • Design of computing systems, ranging from small, embedded elements to large, distributed computing environments • All aspects of the system life cycle, including specification, verification, implementation and synthesis, and testing and evaluation of both hardware and software system components • Principle application area of embedded and real-time systems with special emphasis on the interaction between hardware and software system components University of Kansas A KTEC Center of Excellence 36 Interactive Intelligent Systems • Create intelligent and interactive systems with sufficient intelligence to help humans accomplish important tasks • Multi-modal interfaces to respond intelligently to user requests, process and present large quantities of information in many forms, and to perform tasks with minimal supervision • Artificial intelligence, intelligent agents, information retrieval, data mining, human-computer interaction, modeling, visualization, multimedia systems, and robotics University of Kansas A KTEC Center of Excellence 37 Bioinformatics • Information technology to process, analyze, and present biological data in new, meaningful, and efficient ways • Knowledge discovery and data mining and analysis as they relate to life sciences • Making key advances in bioinformatics methods and tools for genomics and proteomics data analysis and other lifesciences-related problems University of Kansas A KTEC Center of Excellence 38 Radar Systems and Remote Sensing • Radars, microwaves, communications, and remote sensing technologies • New ways to use electromagnetic waves in the remote sensing of the land (surface and subsurface), sea, polar ice, and the atmosphere • Developing new remote sensing sensors (primarily radar), and new methods for solving electromagnetic problems University of Kansas A KTEC Center of Excellence 39 FastTrack Ph.D. • Enter the Ph.D. program directly from the B.S. • Finish in 5 years • 42 course credit hours past B.S. • Possible schedule: Semester 1 Semester 2 3 courses 3 courses Semester 3 Semesters 4-10 2-3 courses + research 0-2 courses + research University of Kansas A KTEC Center of Excellence 40 Deadlines • The application deadline is March 1st, but for full consideration for fellowships and research/teaching assistantships, applications should be received by January 1st. • For more details about the application process please see our graduate admissions page. University of Kansas A KTEC Center of Excellence 41 Websites • • • • www.ittc.ku.edu www.cresis.ku.edu www.eecs.ku.edu www.ku.edu University of Kansas A KTEC Center of Excellence 42