1. Multi-Layer Analysis of Web Browsing Performance for Wireless PDAs Adesola Omotayo & Carey Williamson June 1, 2015

2. 2 Presentation Outline Introduction & Motivation Related Work Data Gathering & Validation HTTP-level Analysis TCP-level Analysis MAC-level & Error Analysis Summary Future Work Introduction & Motivation Related Work Data Gathering & Validation HTTP-level Analysis TCP-level Analysis MAC-level & Error Analysis Summary Future Work

3. 3 Introduction & Motivation Widespread availability of WiFi hot spots Limited understanding of multi-layer protocol interactions over IEEE 802.11b WLAN Crucial to understand the performance of the wireless Web Widespread availability of WiFi hot spots Limited understanding of multi-layer protocol interactions over IEEE 802.11b WLAN Crucial to understand the performance of the wireless Web

4. 4 Related Work Workload of clients at wireline networks Client-based “Changes in Web Client Access Patterns”, P. Barford, A. Bestavros, A. Bradley, and M. Crovella, 1999 Server-based “Internet Web Servers: Workload Characterization and Performance Implications”, M. Arlitt and C. Williamson, October 1997 Proxy-based “On the Scale and Performance of Cooperative Web Proxy Caching”, A. Wolman, G. Voelker, N. Sharma, N. Cardwell, A. Karlin, and H. Levy, December 1999 Workload of wireless clients Local-area “Analysis of a Local-Area Wireless Network”, D. Tang and M. Baker, August 2000 Campus-area “Analysis of a Campus-Wide Wireless Network”, D. Kotz and K. Essien, September 2002 Metropolitan-area “Analysis of a Metropolitan-Area Wireless Network”, D. Tang and M. Baker, August 1999 Workload of clients at wireline networks Client-based “Changes in Web Client Access Patterns”, P. Barford, A. Bestavros, A. Bradley, and M. Crovella, 1999 Server-based “Internet Web Servers: Workload Characterization and Performance Implications”, M. Arlitt and C. Williamson, October 1997 Proxy-based “On the Scale and Performance of Cooperative Web Proxy Caching”, A. Wolman, G. Voelker, N. Sharma, N. Cardwell, A. Karlin, and H. Levy, December 1999 Workload of wireless clients Local-area “Analysis of a Local-Area Wireless Network”, D. Tang and M. Baker, August 2000 Campus-area “Analysis of a Campus-Wide Wireless Network”, D. Kotz and K. Essien, September 2002 Metropolitan-area “Analysis of a Metropolitan-Area Wireless Network”, D. Tang and M. Baker, August 1999

5. 5 Data Gathering & Validation Selected websites news, yellow pages, driving directions, stock quotes, educational resources, and downloadable PDA software Over a period of 35 minutes 398 TCP connections 1.8% with expected FIN handshake 96.5% used the RST packet 1.7% unsuccessful connections Selected websites news, yellow pages, driving directions, stock quotes, educational resources, and downloadable PDA software Over a period of 35 minutes 398 TCP connections 1.8% with expected FIN handshake 96.5% used the RST packet 1.7% unsuccessful connections Wireless Client Internet Wired Network Access Point Wireless Sniffer A very simple workload AP: Netgear WAB 102 PDA: Compaq iPAQ 3600 Pocket PC, Windows CE, IE, MTU size of 1500 bytes Wireless Sniffer: Sniffer Pro 4.60.01, microsecond resolution timestamps

6. 6 HTTP-level Analysis Server Response Time distinct plateaus consistent server response time response times < 200 ms distinct plateaus consistent server response time response times < 200 ms Network RTT dominates the response latency Cache per-destination state information Network RTT dominates the response latency Cache per-destination state information

7. 7 HTTP-level Analysis Web Object Sizes object sizes: 90% < 10 KB 2.5% > 40 KB file types: most prevalent: GIF, JPG & HTML Least prevalent: PNG largest objects transferred: executables object sizes: 90% < 10 KB 2.5% > 40 KB file types: most prevalent: GIF, JPG & HTML Least prevalent: PNG largest objects transferred: executables Cache contents from wireless portals on Proxy Servers Increase support for PNG file type across browsers Compress executable files to be more compact Cache contents from wireless portals on Proxy Servers Increase support for PNG file type across browsers Compress executable files to be more compact

8. 8 HTTP-level Analysis HTTP Transfer Time HTTP transfers 96% < 1 second 2.5% > 2 seconds larger objects take longer to download few small objects have excessively long transfer times HTTP transfers 96% < 1 second 2.5% > 2 seconds larger objects take longer to download few small objects have excessively long transfer times HTTP transfer times are generally low Most responses fit in a single TCP packet HTTP transfer times are generally low Most responses fit in a single TCP packet

9. 9 TCP-level Analysis TCP Connection Type 13% were persistent 87% were non-persistent 4% of TCP connections sent > 10 HTTP requests 65% of HTTP transfers occurred on persistent connections As much as 73 HTTP requests were seen per connection 13% were persistent 87% were non-persistent 4% of TCP connections sent > 10 HTTP requests 65% of HTTP transfers occurred on persistent connections As much as 73 HTTP requests were seen per connection Use persistent connections for all web sites

10. 10 TCP-level Analysis TCP Connection Duration 75% sent < 20 packets 6% sent > 100 packets 80% sent < 10 KB 8% sent > 50 KB 75% lasted < 1 second 10% lasted > 30 seconds 4 connections lasted > 300 sec. 75% sent < 20 packets 6% sent > 100 packets 80% sent < 10 KB 8% sent > 50 KB 75% lasted < 1 second 10% lasted > 30 seconds 4 connections lasted > 300 sec. Most TCP connections are non-persistent Most web object transfers are small Tightly set the persistent connection timeout Most TCP connections are non-persistent Most web object transfers are small Tightly set the persistent connection timeout

11. 11 TCP-level Analysis TCP Connection Throughput Connection Throughput in bits per second (bps) 14 12 10 8 6 4 2 0 0 200000 400000 600000 800000 1e+06 1.2e+06 Distribution of TCP Connection Throughput Frequency in Percent 95% < 400 Kbps Non-persistent TCP connections Small HTTP transfer size Non-negligible RTTs TCP slow start effects Non-persistent TCP connections Small HTTP transfer size Non-negligible RTTs TCP slow start effects

12. 12 MAC-level & Error Analysis MAC-level Retransmissions 3% of the packets 40% of the connections most retry attempts for a packet: 6 3% of the packets 40% of the connections most retry attempts for a packet: 6 CRC Errors 0.04% of the packets TCP-level Retransmissions 0.2% of the packets 12 TCP connections 2 connection have > 3 packet loss 0.2% of the packets 12 TCP connections 2 connection have > 3 packet loss HTTP-level Errors Unsuccessful: 1% Successful: 96.74% Aborted: 2.26% Unsuccessful: 1% Successful: 96.74% Aborted: 2.26% Wireless channel quality does not have a major impact on wireless Web browsing performance

13. 13 Summary (1 of 2) FactsImplications Network RTT dominates the response latency Caching per-destination state information (e.g., RTT, cwnd) might be effective Web objects are typically small Web proxy caching of content from wireless portals could reduce network latency Largest web objects transferred were executables Software providers should compress executable files into more compact file formats Even though free, the least prevalent graphics file type on the web is PNG Increase support for PNG file type across web browsers

14. 14 FactsImplications 87% were non-persistent and 65% of HTTP transfers occurred on persistent connections Wireless Web browsing would be faster if persistent connections were used for all Web sites Some TCP connections lasted longer than 300 seconds Persistent connection timeout should be tightly set 52% of the TCP packets were transmitted by the client PDA Some form of ACK consolidation in Windows CE would economize on wireless network usage and battery power for wireless device MAC: 3% of the packets CRC: 0.04% of the packets TCP: 0.2% of the packets HTTP: 1% of the connections Wireless channel quality does not have a major impact on wireless Web browsing performance Summary (2 of 2)

15. 15 Future Work Expand the work to a large scale traffic measurement Study the effect of interference and range overlapping among closely located APs Expand the work to a large scale traffic measurement Study the effect of interference and range overlapping among closely located APs

16. 16 References M. Arlitt and C. Williamson, “Internet Web Servers: Workload Characterization and Performance Implications”, IEEE/ACM Transactions on Networking, Vol. 5, No. 5, pp. 631-645, October 1997. P. Barford, A. Bestavros, A. Bradley, and M. Crovella, “Changes in Web Client Access Patterns”, World Wide Web Journal, 1999. D. Kotz and K. Essien, “Analysis of a Campus-Wide Wireless Network”, Proceedings of ACM MOBICOM, Atlanta, GA, pp. 107-118, September 2002. D. Tang and M. Baker, “Analysis of a Metropolitan-Area Wireless Network”, Proceedings of ACM MOBICOM, Seattle, WA, pp. 13-23, August 1999. D. Tang and M. Baker, “Analysis of a Local-Area Wireless Network”, Proceedings of ACM MOBICOM, Boston, MA, pp. 1-10, August 2000. A. Wolman, G. Voelker, N. Sharma, N. Cardwell, A. Karlin, and H. Levy, “On the Scale and Performance of Cooperative Web Proxy Caching”, Proceedings of ACM SOSP, December 1999. M. Arlitt and C. Williamson, “Internet Web Servers: Workload Characterization and Performance Implications”, IEEE/ACM Transactions on Networking, Vol. 5, No. 5, pp. 631-645, October 1997. P. Barford, A. Bestavros, A. Bradley, and M. Crovella, “Changes in Web Client Access Patterns”, World Wide Web Journal, 1999. D. Kotz and K. Essien, “Analysis of a Campus-Wide Wireless Network”, Proceedings of ACM MOBICOM, Atlanta, GA, pp. 107-118, September 2002. D. Tang and M. Baker, “Analysis of a Metropolitan-Area Wireless Network”, Proceedings of ACM MOBICOM, Seattle, WA, pp. 13-23, August 1999. D. Tang and M. Baker, “Analysis of a Local-Area Wireless Network”, Proceedings of ACM MOBICOM, Boston, MA, pp. 1-10, August 2000. A. Wolman, G. Voelker, N. Sharma, N. Cardwell, A. Karlin, and H. Levy, “On the Scale and Performance of Cooperative Web Proxy Caching”, Proceedings of ACM SOSP, December 1999.

17. 17 Thank You! ? ?