Large Audience Environments Any location with a large WiFi user population WBA projected a growth rate of 350% per year for such WiFi deployments Various successful deployment models such as Boingo-Google, Mobily-Aruba etc. already exist Source: WBA(Wireless Broadband Alliance) Source: Wireless Broadband Alliance (WBA)
What about User Experience ? As number of active clients increases, user experience diminishes significantly
Well Known Factors Contention and collision – Increases with growing competition Rate Diversity – Slower STA slows down all other STAs – Various fairness realizations like WFQ, TBR etc. Random Losses and TCP performance – TCP treats packet losses as congestion signals – Usage of TCP ECN and proxy servers isolate wired and wireless networks Traffic Asymmetry
Downlink Traffic dominates for 90% of data traces
Traffic Asymmetry Majority of data traffic is Web based Downlink Traffic dominates compared to uplink Majority of data packets are for HTTP based web activities
Traffic Asymmetry In scope of our problem it is: – Downlink/Uplink Asymmetry – Channel Access Asymmetry Implications: – Packets spend more time at AP’s TxQ – Frequent packet drops
Wireless Channel Uplink Traffic Downlink Traffic Channel access for uplink traffic is more
Priority Control Packet A ACK DIFS DIFS H Channel Access N Slots Busy medium Wins Contention STA C STA B STA A Smaller IFS ClassCWminCWmaxAIFSTXOP Limit AP15164 STAs510N/A
Linear Scaling Priority Model Priority Level Linear relationship between Goodput and Priority Level
Adaptive Prioritization 100 ms DHDHDDDHDD Time High Priority Default Priority Decision Points Priority Level 3
Test Bed 2600 Sqft Area with multiple APs, 45 STAs Netgear b/g wireless cards with Atheros chipsets and MADWIFI drivers Latency emulation using DummyNet Modified SURGE for web traffic generation Requests inter-arrival closely follows the ones observed for SIGCOMM traces Uplink UDP traffic using Iperf to emulate Background Traffic
Performance Experiment involves sending 25 requests and observe response for 4 minutes duration Request Serving rate is 4 times better than NPC Experiment involves sending 25 requests and observe response for 4 minutes duration Request Serving rate is 4 times better than NPC
Robustness Performance in presence of Multiple Aps ?? WiFox w/o WiFox
Robustness w/o WiFox WiFox Unfair Distribution Fairness Realization ??
Performance: TxQ Dynamics WiFox w/o WiFox
Conclusion WiFox Delivers: – Deployment Ready Solution – Enhanced user experience with % Downlink Goodput improvements 40-60% faster response time Open Problems: – Characterizing asymmetry problem for n – Support for real time applications like chats etc. – QoS
30 Merci !!
Multi AP Scenario DDDDDD DDDDDDD AP 1 ( Priority Level 4) AP 2 ( Priority Level 3) DDDDDDD DDDDD AP 1 ( Priority Level 3) AP 2 ( Priority Level 5) time 100 ms
Performance: Insight Enables AP to switch to HIGH priority state under heavy load Avoids TxQ saturation Significant reduction in ReTx rate compared to stock WiFi implementation
Robustness: Uplink Traffic Scenarios where few users indulge in heavy uplink activities like video uploading, cloud synchronization etc.