Why it Takes so Long to Connect to a WiFi Access Point
Changhua Pei, Zhi Wang, Youjian Zhao, Zihan Wang, Yuanquan Peng, Wenliang Tang, Yuan Meng, Dan Pei Xiaodong Qu
INFOCOM 2017 2
Measurement Correlation Motivation Modeling Results Analysis 3
Measurement Correlation Motivation Modeling Results Analysis 4 WiFi is indispensable in our daily lives
Source: Cisco VNI Mobile, 2017
97.4 (49% of overall traffic) 3X 5X 454 30.8 94
2005 2020 2006 2020 Overall WiFi Traffic Growth Hotspots (Exabytes) (Millions) 5 Experience of WiFi Network Remote Services Mobile Device AP
Throughput
Downloading 6 Experience of WiFi Network Remote Services Mobile Device AP
WiFi Hop Latency
Online Gaming 7 Experience of WiFi Network Remote Services Mobile Device AP 8 Experience of WiFi Network Mobile Device AP
I want to connect to the AP! Connection set-up process
Now I can talk to Connection time cost the Internet ^^ Urgent need to study the connection set-up time 9
Mobile Device AP Suranga [WiNTECH’13] is the first work focus on WiFi connection time cost : • The connection set-up process in the wild is unknown • Lack thorough investigation in a larger scale. I want to access theWe AP!focus on: Connection set-up procedure • How about the connection time cost in the wild? Now• WhatI canistalkthetoculprit of theConnectionhigh connectiontime costtime cost? the• WhatInternetcan ^^I do to reduce the connection time cost? 10
Measurement Correlation Motivation Modeling Results Analysis 11 DATASET
• WiFi Manager of Tencent Technology • Provide Free WiFi service • Top in the Android/iOS App market (China) • About 50K downloads every day • Continuously collect one week data from May 3 to May 9. 12 DATASET
Connection Log Dataset
l 7 Million unique APs l 5 Million unique mobile devices l 4 different cities. l 0.4 billion overall connection attempts. 13 CDF of the connection time cost
100 80 60 40 CDF (%) CDF 20 0 0 5 10 15 20 25 30 Connection Time Cost (s)
15% (5%) successfulGlobalconnectionsPublic WiFi Hotspotsconsume over 5 (10) seconds! 14 DATASET
Breaking Down Dataset
l 12,472 selected devices l 706K connection attempts l Spread over differentplaces. Connection Time Cost 15 App starts timing
Scan …
Association Association Request 4-way handshake Authentication Request IP DHCP … Obtain IP Connection Time Cost 16 App starts Disconnected timing
Scan …
Association Associating Association Request 4-way Associated 4-way handshake handshake Authentication Request completed IP DHCP … Obtain IP connected 17 WiFi Association: Success vs. Failure 54.9% 60 54.9 50 40 45% 30
20 14.6 9.4 percentage(%) 7.85 10 7.76 5.042 0.377 0.071 0 Success Timeout DHCP Wrong Swtich to Forget WiFi Unkown Failure Password another WiFi Switched WiFi Off
Based on the breaking down dataset. 18
Does there exist one sub-phase which dominates the overall connection set-up process? 100 100
80 80 19 Scan 60 60 40 40 Association CDF (%) CDF 0s~7s (%) CDF 0s~7s 20 7s~15s 20 7s~15s 15s~30s 15s~30s 0 0 0 20 40 60 80 100 0 20 40 60 80 100 Proportion of Scan (%) Proportion of Association (%) 100 100 0s~7s 80 80 7s~15s 15s~30s 60 60 40 CDF (%) CDF 40 DHCP 0s~7s (%) CDF 20 Authentication 7s~15s 20 15s~30s 0 0 0 20 40 60 80 100 0 20 40 60 80 100 Proportion of Authentication (%) Proportion of DHCP (%) 100
80 20 Association and Authentication 60 40 Association do not take too much time. (%) CDF 0s~7s 20 7s~15s 15s~30s 0 0 20 40 60 80 100 Proportion of Association (%) 100
80 60 Fixed number of MAC-layer 40 CDF (%) CDF 0s~7s packets exchange in Association Authentication 20 7s~15s or Authentication. 15s~30s 0 0 20 40 60 80 100 Proportion of Authentication (%) 100
80 21 Scan 60 0s-7s, 7s-15s: DHCP phase 40 occupies more than 80%, which CDF (%) CDF 0s~7s 20 7s~15s is consistent with WiNTECH work. 15s~30s 0 0 20 40 60 80 100 Proportion of Scan (%)
100 0s~7s 80 7s~15s 15s~30s 15-30s: Scan phase 60 40 DHCP consumes more than (%) CDF DHCP phase for those 20 0 > 15s. 0 20 40 60 80 100 Proportion of DHCP (%) 22 Scan Phase dominate the connection
0s~7s 7s~15s 15s~30s 100 0s-7s class: Scan consume 80 less than 3.4s. 60
40 11.6s CDF (%) CDF 20 3.4s 15s-30s class: For more than 0 40% processes, Scan phase 0 5 10 15 20 25 30 consume more than 11.6s. Scan Time Cost (s)
• Why does this happen? 23 Anomalous transitions cause long scanning 290 • Anomalous transition to 2415 578 Scanning Associating Disconnectedstate
1798 3066 1576 • Mobility 1229 669 • WiFi interference 137 • 432 Disconnected Associated System process delay • …
9 352 1661 1309 Connected Completed Total number of transitions in the dataset. 24 Take-away messages: 1. For those connection whose time cost > 15s, Scan is the dominate sub-phase. 2. Scan dominates the whole process because there are anomalous transitions. 25
Measurement Correlation Motivation Modeling Results Analysis 26 Which feature affect the connection time cost the most? 1. Give Intermediate results to gain some intuitions before the ML model. 2. Help feature selection. 27 Introduction of the Connection Log Dataset Aggregated results 28 Visualization analysis of all the APs. 10
9
8
7
6
Connection Time Cost (s) Cost Time Connection 5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Hour of Day • Association timing affects the connection time cost. 29 Visualization analysis
6 5 • Connection with higher RSSI 4 tend to have smaller average 3 2 connection time costs. 1 Connection Time Cost (s) Cost Time Connection -95 -90 -85 -80 -75 -70 -65 -60 RSSI (dBm) 30 Visualization analysis
6 • The larger the number of 5 4 associated devices is, the 3 2 higher average connection 1 Connection Time Cost (s) Cost Time Connection 0 40 80 120 160 200 240 time cost. Number of Devices 31 Correlation Analysis v Kendall correlation: (rank correlation)
���������� ����� − ���������� ����� ��� = �(� − 1)/2 v Relative Information Gain: (RIG) how much a factor X helps to predict the final latency Y