Lessons learnt from the Beijing Olympic Games

Transcription

1 Lessons learnt from the Beijing Olympic Games Website Measurement Rocky K. C. Chang The Internet Infrastructure and Security Group The Hong Kong Polytechnic University October 28 1

2 Motivations Test and compare our measurement system (OneProbe) with other methods. Compare and evaluate the web performance at different sites for this global event. Discover new and interesting results about Internet path and web performance. 2

3 Outline Measurement methodology Measurement results Overall results A measurement tool comparison Correlation of loss and delay peak Asymmetric loss patterns Effects of network configuration changes Conclusions and future works 3

4 Measurement methodology Uncooperative methods Ping g( (ICMP) PPing (TCP SYN-ACK) HTTPing (TCP SYN-ACK and HTTP data) OneProbe (TCP data) Traceroute on forward path Changes in TTL on reverse path Measuring points A data center in Hong Kong Three Beijing origin i servers, three Chinacache h sites, and three Akamai/Quest sites 4

5 The three sets of network paths Three Chinacache sites One of the went through the Korea Network Information Center. They all went through the CNC Group Backbone. The servers were located in three different provinces: Henan, Hebei, and Shandong. Three Beijing sites The paths were very similar il (going through h the same subnets). Three Akamai/Quest sites The first four of the five hops were the same. The sites were located in Hong Kong. 5

6 Measurement parameters Measurement period During the entire Olympic Games period and one week after the Games Sampling rate OneProbe: sending a probe every.5 secs Others: sending a probe every sec. Probe packet size OneProbe: 15 bytes Ping: 1 and 15 bytes PPing: 4 bytes HTTPing: variable 6

7 Path metrics Ping, PPing, and HTTPing RTT and round-trip loss rate OneProbe RTT One-way loss rates One-way reordering rates One-way capacity 7

8 Overall path quality Chinacache: RTT: 87ms-26ms Loss rate: <= 1% Beijing RTT: 78ms-132ms Loss rate: <= 2% Akamai/Quest RTT: 3.4ms-4.9ms Loss rate: <= 1% No reordering events observed Mostly stable forward-path routes and reverse-path hop counts Detected configuration changes Some persistent load-balancing 8

9 RTT measurement Diurnal patterns Weekdays and weekends HTTPing s and PPing s results are compatible with OneProbe s. HTTPing s RTTs are slightly higher. ICMP Ping RTTs do not always match with OneProbe s. 9

10 25 Round trip delay observed by different tools 2 Round t trip Delay (ms) OneProbe HTTPing PPing Ping x 1 9 1

11 Packet loss patterns Diurnal patterns Correlation of the RTT peaks with the round-trip ppacket losses Ping s packet loss accuracy Dominance of the reverse-path loss for the Chinacache and Beijing sites 11

12 Distribution of Round trip loss measured by Ping, PPing and OneProbe 1 Distribution of Round trip loss measured by Ping, PPing, and OneProbe CDF.5 CDF OneProbe 1st Packet OneProbe 2nd Packet Ping PPing Loss Rate (%).3.2 OneProbe 1st Packet OneProbe 2nd Packet.1 Ping PPing Loss Rate (%) Distribution of round trip loss measured by Ping, PPing and OneProbe 1 CDF OneProbe 1st Packet.2 OneProbe 2nd Packet.1 Ping PPing Loss Rate (%)

13 Timeseries of Round trip delay and loss rates 3 OneProbeRTT (ms) 2 1 (% %) pingLikeLoss1stPkt (%) (%) fw1stPktLoss 27bw1stPktLoss(All)

14 Ranking of difference between measured forward and backward packet loss 2 Ranking of difference between measured forward and backward packet loss f Percentage Loss s Rate Difference of Percentage Loss s Rate Difference of

15 Special cases Effects of detectable route changes on the path performance Effects of undetectable route changes on the path performance 15

16 Timeseries of RTT, loss rate and route change 3 OneProbeRTT (m ms) 2 1 (%) 2 1 Round trip 1st PktLoss (% %) (% %) (Ho op) 2 Forward 1st Pkt Loss 1 2 Backward 1st Pkt Loss TCPTraceroute x 1 9

17 Conclusions Ping is not reliable for RTT measurement. It could significantly underestimate the delay experienced by TCP data. Ping, PPing, and HTTPing cannot measure beyond round-trip loss rates. E.g., cannot detect highly asymmetric path losses. The protocol used in the probes should match with the protocol under measurement. A careful measurement study can reveal configuration changes on the path. 17

18 Acknowledgments The experiments and analyses were conducted by the following members in the group: Edmond Chan Waiting Fok Daniel Luo This project is partially supported by a grant from the Innovation Technology Fund in Hong Kong. 18