First PlanetLab Asia Workshop Outline Happy Packets to You! • Why this study? • Methodology Randy Bush • Results and analysis Timothy G. Griffin • Open issues Jun Li Z. Morley Mao • Conclusions Eric Purpus Dan Stutsbach Happy Packets to You! 1 9/17/2004 This research is funded by NSF award ANI-0221435. Jun Li, U of O, 09/17/04 Why This Study? Happy Packets • We frequently hear comments about Internet • What ultimately counts is whether the control plane quality, such as customer's packets can reach their intended – Internet routing is fragile and collapsing, destination with good performance – Yesterday was a bad routing day on the Internet, – Namely, the performance at data plane – BGP is broken or is not working well, – Changing protocol X to Y will improve routing, or – And after all, this is the functionality of the – Internet routing has been severely affected by event X control plane (e.g. power blackout, worm outbreak) • We call them happy packets • But what measurement can really tell the quality of control plane? – Number and frequency of BGP updates? Happy Packets to You! 2 9/17/2004 Happy Packets to You! 3 9/17/2004 Our Goal Methodology • Answer this question: Are packets happy • How to measure packet happiness at the under routing changes? data plane? – Basically, we evaluate Internet control plane – Use the PlanetLab quality by measuring the data plane • How to introduce routing changes into the performance control plane – Use a BGP Beacon Happy Packets to You! 4 9/17/2004 Happy Packets to You! 5 9/17/2004 1
Happiness over the PlanetLab • A set of geographically and topologically sink diverse PlanetLab nodes are selected as probe sites • A site from Seattle is selected as a sink • Every probe site continuously sends testing UDP streams toward the sink site – While the routing toward the sink changes • Over a period of four months Happy Packets to You! 6 9/17/2004 Happy Packets to You! 7 9/17/2004 Metrics of Happiness BGP Beacon • Using well-established metrics (DDJ&R): • An unused globally visible IP address prefix – Delay, drop, jitter, and reordering • With a schedule of BGP announcements • Delay: the relative to the mean one-way delay and withdrawals regarding reaching the • Loss rate: % of packets dropped per second prefix • Loss duration: the length of a time window with • We use a multi-homed BGP Beacon exceptionally high loss rate 192.83.230.0/24 • Jitter: delta between delays • The test stream sink has a specific IP • Reordering rate: % of reordered packets per address from this prefix second Happy Packets to You! 8 9/17/2004 Happy Packets to You! 9 9/17/2004 Collecting Control Plane Data • Oregon RouteViews Project archive BGP updates • Can help observe BGP updates related to the BGP Beacon prefix – Thus BGP duration and BGP update number during an event can be measured Happy Packets to You! 10 9/17/2004 Happy Packets to You! 11 9/17/2004 2
What Did We Find? Results of an Average Stream • Stream from 128.95.219.192 as an example • Average and worst case DDJ&R results of • Under four different routing changes over 20 min: individual streams AB-B, AB-A, A-AB, B-AB • Aggregated results of DDJ&R • Performed well in general, either during or outside • Control plane data routing changes – Most times packet delays are acceptable – No reordering was detected – Thus jitter is also acceptable – A 30-sec loss duration in the AB-A case Happy Packets to You! 12 9/17/2004 Happy Packets to You! 13 9/17/2004 Happy Packets to You! 14 9/17/2004 Happy Packets to You! 15 9/17/2004 Happy Packets to You! 16 9/17/2004 Happy Packets to You! 17 9/17/2004 3
Results of a Worst-Case Stream • Stream from lcs-bgp.vineyard.net • Under four different routing changes over 20 min: AB-B, AB-A, A-AB, B-AB • Performed the worst compared to other streams – Longer delay than others – Longest loss duration • 10s in AB-B w/ 91 drops & 8 reorders • However, not significantly worse than its own normal period Happy Packets to You! 18 9/17/2004 Happy Packets to You! 19 9/17/2004 Happy Packets to You! 20 9/17/2004 Happy Packets to You! 21 9/17/2004 Aggregated Results • Delay CDF • Jitter CDF • Loss rate • Reordering – not plotted (close to 0) Happy Packets to You! 22 9/17/2004 Happy Packets to You! 23 9/17/2004 4
CDF of (Relative) Delay Jitter CDF Results for Results for AB-B, AB-B, A-AB, A-AB, B-AB B-AB are similar are similar Happy Packets to You! 24 9/17/2004 Happy Packets to You! 25 9/17/2004 Loss Rate Loss Rate Happy Packets to You! 26 9/17/2004 Happy Packets to You! 27 9/17/2004 Misconception in Inferring DDJ&R Summary Packet Happiness • Acceptable during injected routing changes • Can control plane data, such as those from – Although generally worse than normal periods RouterViews or RIPE, predict packet • In most cases, BGP performs well performance, thus equivalent to measuring • Can also approximate closely the packet delivery DDJ&R? performance between two routers • Answer: No! – DDJ&R of each UDP stream from end to end is also close to the BGP Beacon router and the router for the probe site Happy Packets to You! 28 9/17/2004 Happy Packets to You! 29 9/17/2004 5
Loss duration vs. BGP update duration Loss duration vs. BGP update number Happy Packets to You! 30 9/17/2004 Happy Packets to You! 31 9/17/2004 White Blood Cells Summary • Perhaps BGP announcements are like white • No clear correlation between loss duration and BGP duration, or loss duration and blood cells number of BGP updates • Their presence may signal a problem • RouteViews archives only provide partial • But they are often part of the cure, not knowledge of the control plane necessarily part of the problem • One should be cautious in using BGP updates to analyze control plane quality Happy Packets to You! 32 9/17/2004 Happy Packets to You! 33 9/17/2004 Open Issues Conclusions • Large-scale control plane events • Data plane performance is the best measure of control plane effectiveness • Congestion effects on DDJ&R • Not only for BGP, but also other routing protocols • Usage of partial control plane knowledge • We’ve found little proof that BGP is not resilient • . . . . or performing poorly during routing changes • And we should be critical about using partial control plane data for study Happy Packets to You! 34 9/17/2004 Happy Packets to You! 35 9/17/2004 6
Questions? Randy Bush <randy@psg.com> Timothy G. Griffin <tim.griffin@intel.com> Jun Li <lijun@cs.uoregon.edu> Z. Morley Mao <zmao@eecs.umich.edu> Eric Purpus <epurpus@cs.uoregon.edu> Dan Stutsbach <agthorr@cs.uoregon.edu> 7
Recommend
More recommend
