WEB traffic characterization and performance analysis

Transcription

1 WEB traffic characterization and performance analysis

2 Summary Objectives of our work Web traffic characterisation through measurements Performance evaluation and analytical model validation Packet level analysis Conclusions 2

3 Objectives of our work Modelling the behaviour of Web clients Page size, page request frequency Single client, aggregate of clients Models and criteria for link dimensioning Effect of single/multiple bottleneck on performances Simple models for performance evaluation Models and criteria for buffer dimensioning Analysis of packet level dynamics 3

4 Data collection WS Internal Traces Drogo ( proxy ) CSELT Network GW Internet External Traces 100 Mbit / sec Client Server TCPDUMP Packet Traces TCP-REDUCE List of TCP connections other 8080 && dest. = drogo Internal Web Connections External Web Connections www && Source = drogo 4

5 WEB pages reconstruction Pages downloading process (HTTP/1.0) Pages reconstruction through TCP connections temporal clustering TCP Connections List HTML- REDUCE WWW level Stats 5

6 Main results Mean Std_dev Distribution Page size LogNormal bytes bytes Page duration Sec Sec. LogNormal Objects per LogNormal + δ page Think Time Sec Sec. LogNormal Pages Inter Req. Time Sec Sec. Exponential Active clients: 9% 15% 6

7 Page size distribution 1.E+00 1-CDF 1.E-01 1.E-02 EMPIRICAL LOGNORMAL WEIBULL o u t l i a r s 1.E-03 1.E-04 1.E-05 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 PAGE SIZE [BYTES] Doubly-Logarithmic Scale 7

8 Think-Time distribution 1.E+00 EMPIRICAL LOGNORMAL 1.E-01 1-CDF 1.E-02 1.E-03 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 THINK TIME [SEC] Doubly-Logarithmic Scale 8

9 Inter-Pages request time 1.E+00 1.E-01 EMPIRICAL 1-CDF 1.E-02 1.E-03 WEIBULL. EXPONENTIAL 1.E-04 1.E-05 1.E-02 1.E-01 1.E+00 1.E+01 IPT [SEC] Doubly-Logarithmic Scale 9

10 Inter-arrival based model Driven by Page Size Inter Request Time Page request frequency is load-independent Cli 1 Cli 2 Cli 3 10

11 Think-Time based model Driven by Page size Think Time Frequency of page requests is load-dependent Cli 1 Cli 2 Cli 3 11

12 Single bottleneck scenario (1) Clients located on high speed LAN Router and server latency is negligible Infinite buffer space: no lost packets 12,500,000 bytes/sec 73,750 bytes/sec 12,500,000 bytes/sec Page 2 Page 1 C M/G/1/PS Model 12

13 Single bottleneck scenario (2) 25 Average Downloading Time vs offered load [sec] ,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 ρ 13

14 Multiple bottlenecks scenario Clients attached to low speed lines Routers and Servers latency is negligible Infinite buffer space 10 Mbit/sec Server addizionale Server addizionale 56.6 Kbit/sec Client osservati 10 Mbit/sec Traffico osservato Router1 Router2 Router Kbit/sec Server osservati Client addizionali Client addizionali 14

15 Two bottlenecks scenario Average downloading time vs offered load Download Time [sec] 8 6 MediaResponseTime MODEM MG1PS MODEM+MG1PS PAB ρ [%] 15

16 Three bottlenecks scenario Download Time [sec] MediaResponseTime MODEM MG1PS-link1 MG1PS-link2 MODEM+MG1PS1 MODEM+MG1PS1+MG1PS2 PAB ρ [%] 16

17 Packet level analysis Real traces are self similar in the range [few msec. - tens of sec.] with H 0.8 Simulated traces shows similar characteristics in the range [10 msec sec.] with H 0.73 unless the traffic is very high (> 0.8) Self similarity has to be taken into account for buffer dimensioning even though it is irrelevant for link dimensioning 17

18 Conclusions and future work The M/G/1/PS model provides good estimation of the average transmission time With multiple bottlenecks, present results indicate that the delay is additive Further analysis are required to confirm this result Buffer occupancy is overestimated by the M/G/1/PS model: models at the packet level taking into account self similarity are required 18