wifi can be the weakest link of round trip network
play

WiFi Can Be the Weakest Link of Round Trip Network Latency in the - PowerPoint PPT Presentation

WiFi Can Be the Weakest Link of Round Trip Network Latency in the Wild Changhua Pei , Youjian Zhao, Guo Chen, Ruming Tang, Yuan Meng, MinghuaMa, Ken Ling, Dan Pei Tsinghua University Carnegie Mellon University 1


  1. WiFi Can Be the Weakest Link of Round Trip Network Latency in the Wild Changhua Pei †, Youjian Zhao†, Guo Chen†, Ruming Tang†, Yuan Meng†, MinghuaMa†, Ken Ling‡, Dan Pei† †Tsinghua University ‡Carnegie Mellon University 1

  2. WiFi is indispensable in our daily lives v Overall WiFi Traffic Growth Source: Cisco VNI Mobile, 2016 2

  3. WiFi is indispensable in our daily lives! v Booming of the Access Points: Number of Access Points! Source: Maravedis, Cisco VNI Mobile, 2016 3

  4. WiFi performance is far from satisfactory Unsatisfactory 100 25ms 80 CDF(%) 60 40 wired part 20 wireless part 0 0.1 1 10 100 1000 10000 RTT (ms) Stringent Threshold: 20~30ms 4

  5. ABANDON THE PAGES USERS WILL USERS EXPECT LEADS TO WiFi performance is far from satisfactory PAGE LOAD TIME > 3 SECONDS 47% 40% PAGE LOAD TIME < 2 SECONDS Akamai study. http://goo.gl/2pwozG. 5

  6. WiFi performance is far from satisfactory 10 1000 ms ms LAST-MILE DELAY PAGE LOAD TIME increase increase Bismark Paper: S. Sundaresan, N. Feamster, R. Teixeira, N. Magharei, et al. Measur- ing and mitigating web performance bottlenecks in broadband access 6 networks. In ACM Internet Measurement Conference , 2013.

  7. WiFi performance is far from satisfactory Stringent Threshold: 20~30ms 7

  8. Challenge: Large Search Space of AP parameters Channel Transmit Location? Channel? Width? Power? 1 BLIND SEARCH among all re- configuration possibilities 6 11 Don’t know the effect before the re- configuration 8

  9. Configurable WiFi Hop WiFi Factors Parameters Latency 1. How to accurately measure the WiFi hop latency ? Retry Transmit Ratio Power? Airtime Utilization Throughput Location? Domain DELAY Gap Model Knowledge RSSI Channel? Physical 2. How to predictthe WiFi hop latency usingWiFi factors Rate effectively? Queuing Channel Length Width? 3. How to use this model to help AP owners to tune their APs? 10

  10. Measurement Training Trace Problematic ML Model Transmit AP Power? Optimization Reconfigure Location? which ? WiFi Channel? Factors Channel for this AP Width? 12

  11. Measurement Training Trace Problematic ML Model Transmit AP Power? Optimization Reconfigure Location? which ? WiFi Channel? Factors Channel for this AP Width? 14

  12. Measuring WiFi Hop Latency: Background AP Server Client TCP SYN UL PL S WL+S RTT PL TCP SYN-ACK DL TCP ACK 12

  13. Measuring WiFi Hop Latency: existing approaches need client-side involvement AP Server Client v RTT : Using PING at client side: RTT = t 3 -t 0 TCP SYN UL PL client-side assistance S WL+S RTT PL TCP SYN-ACK DL TCP ACK 13

  14. Measuring WiFi Hop Latency: existing approaches need client-side involvement AP Server Client v RTT : Using PING at client side: RTT = t 3 -t 0 TCP SYN UL PL client-side assistance S WL+S RTT PL v DL : Packet Capture: TCP SYN-ACK DL DL = t 3 – t 2 ’ TCP ACK Time synchronization 14

  15. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake PL packets WL+S RTT PL TCP SYN-ACK DL TCP ACK 15

  16. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake DL PL packets WL+S RTT PL TCP SYN-ACK DL MAC layer ACK TCP ACK 16

  17. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake DL t 3 ’-t 2 ’ √ PL packets WL+S RTT PL TCP SYN-ACK DL MAC layer ACK TCP ACK 17

  18. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake DL t 3 ’-t 2 ’ √ PL packets WL+S UL t 4 ’-t 3 ’ √ RTT PL TCP SYN-ACK DL MAC layer ACK TCP ACK 18

  19. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake DL t 3 ’-t 2 ’ √ PL packets WL+S UL t 4 ’-t 3 ’ √ RTT PL Data TCP SYN-ACK DL packets DL t 3 ’-t 2 ’ √ MAC layer ACK TCP ACK 19

  20. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake DL t 3 ’-t 2 ’ √ PL packets WL+S UL t 4 ’-t 3 ’ √ RTT PL Data TCP SYN-ACK DL packets DL t 3 ’-t 2 ’ √ MAC layer ACK UL delay-ACK � TCP ACK 20

  21. Measuring WiFi Hop Latency: all measurements on APs AP Client Server Delay Description Type 3-way WL t 2 ’-t 1 ’ √ TCP SYN UL handshake DL t 3 ’-t 2 ’ √ PL packets WL+S UL t 4 ’-t 3 ’ √ RTT PL Data WL S � TCP SYN-ACK DL packets DL t 3 ’-t 2 ’ √ MAC layer ACK UL delay-ACK � TCP ACK Use the latest 3-way handshake packet to approximate data packets’ WL and UL! 21

  22. Data collection v Real deployment in Tsinghua University in China. v 47 free Netgear WNDR4300 router equipped with Openwrt v 44 in dormitory, 3 in department of computer science v Continuously collected from May 20 th to July 20 th v Collected about 2 terabytes raw data trace 22

  23. 10% packets’ Measurement Result WiFi hop latency >100ms 50% packets’ WiFi hop latency >20ms 23

  24. Measurement Result For nearly 50% of the domestic packet, over 60% of the time is occupied by WiFi hop delay. 24

  25. Measurement Training Trace Problematic ML Model Transmit AP Power? Optimization Reconfigure Location? which ? WiFi Channel? Factors Channel for this AP Width? 28

  26. Predicting the Latency using WiFi factors WiFi Hop Latency WiFi Factors (Fast vs. Slow) as labels as features Machine Learning Predicting Model 26

  27. Abbr. WiFi factors Description Generated By AU airtime utilization % of channel time used by all the traffic iw info Q queue length Number of packets queued in hardware queue. debugfs snapshot RR retry ratio %packets retried in IEEE 802.11 MAC-layer. iw info RSSI RSSI Received signal strength of UE associated on AP. iw info T tx transmitting Bytes sent to UE every 10s. ifconfig info throughput T rx receiving Bytes received from UE every 10s. ifconfig info throughput RPR receiving physical Snapshot of physical rate for receiving packets iw info rate from UE. TPR transmitting Snapshot of physical rate for sending packets to UE. iw info physical rate 27

  28. Visualization and Correlation analysis Purposes: • Intermediate results to gain some intuitions • Help explain the ML results. 28

  29. Visualization of the correlation Negative Trends No Clear Trends Positive Trends Airtime Utilization Transmitting Physical Rate Receiving Throughput Transmitting Throughput Retry Ratio RSSI Queue Snapshot Receiving Physical Rate 29

  30. Visualization of the correlation Negative Trends No Clear Trends Positive Trends Airtime Utilization Transmitting Physical Rate Receiving Throughput Transmitting Throughput Retry Ratio RSSI The model is general because almost No strong effect on WiFi hop latency when : all parameter spaces are covered AU < 0.5 or TPR > 60 Mbps or RSSI > -60 dbm Queue Snapshot Receiving Physical Rate thanks to the variety of the data. 30

  31. Quality Kendall RIG Correlation Analysis Metric AU 0.86 0.05 v Kendall correlation: (Kendall) RSSI -0.5 0.06 !"# = &'(&')*"(! +",)- − *,-&')*"(! +",)- RR 0.4 0.08 ((( − 1)/2 TPR -0.3 0.11 RPR -0.2 0.09 v Relative Information Gain: (RIG) T rx -0.17 0.01 how much a factor helps to Q 0.15 0.007 predict the final latency T tx -0.006 0.02 31

  32. Quality Kendall RIG Correlation Analysis Metric AU 0.86 0.05 v TPR is the best choice to RSSI -0.5 0.06 present the latency. This is RR 0.4 0.08 because of the rate adaption TPR -0.3 0.11 algorithm. RPR -0.2 0.09 T rx -0.17 0.01 Q 0.15 0.007 T tx -0.006 0.02 32

  33. Decision Tree ( AU, RR, RSSI, T rx ,T tx , TPR, RPR) SLOW/FAST Decision Tree Predicting Model 33

  34. Decision Tree v 4 FAST: :;, =; < 12.5 A-, :; + =; < 25 A- SLOW::;, =; ≥ 12.5 A-, :; + =; ≥ 25 A- v Package: scikit learn package v Evaluation: 10-fold validation 34

  35. Decision Tree Method Latency Type Accuracy Truth False Positive Positive Rate Rate Decision DL 0.78 0.76 0.24 Tree UL 0.68 0.67 0.27 DL+UL 0.77 0.79 0.31 35

  36. Decision Tree Method Latency Type Accuracy Truth False Positive Positive Rate Rate Decision DL 0.78 0.76 0.24 Tree UL 0.68 0.67 0.27 DL+UL 0.77 0.79 0.31 v The Random Forest, ( tree number = 200, tree depth = 100), Accuracy > 0.8 with 0.21 False Positive Rate for DL. v Why Decision Tree instead of Random Forest? interpretability+ usability 36

  37. Decision Tree 37

  38. Measurement Training Trace Problematic ML Model Transmit AP Power? Optimization Reconfigure Location? which ? WiFi Channel? Factors Channel for this AP Width? 38

Recommend


More recommend