Mitigating the Impact of Physical Layer Capture and ACK Interference in Wireless Networks Wang Wei

School of Computing WiFi is Ubiquitous

School of Computing The Trends of WiFi Higher density WiFi hotspot market: Annual growth at 84% Global WiFi Hotspot Market , by Research and Markets More traffic WiFi WiFi will make up 56% of Internet traffic in 2017 Cisco Visual Networking Index forecast,

School of Computing The Problems of WiFi Higher density More traffic Serious interference

School of Computing The Problems of WiFi Serious interference Unfairness due to capture Pitfall of Message in Message Mechanism MAC ACK interference

School of Computing Thesis Outline Unfairness due to capture Pitfall of Message in Message Mechanism MAC ACK interference FairMesh Adaptive MIM MinPACK

School of Computing Link Layer Unfairness due to Capture Common in mesh networks, esp. near Gateway Some nodes have little chance to TX successfully Seriously affect user experience

School of Computing Link Layer Unfairness due to Capture Common in mesh networks, esp. near Gateway A GW N B C E F D X Part of the network is un-usable!

School of Computing Design Challenges The unfairness problem is complex Multiple causes for unfairness Arbitrary topology/traffic Tradeoff between fairness and efficiency No central controller in mesh networks

School of Computing Three Canonical Scenarios Asymmetric topology Direct capture Indirect capture 1 dB difference of RSSI can result in capture effect! ABCDABC ABCD

School of Computing Three Canonical Scenarios Asymmetric topology Direct capture Indirect capture ABCDABC ABCD Offender Victim

School of Computing Degree of Unfairness ABC

School of Computing Degree of Unfairness ABC

School of Computing Degree of Unfairness ABC AB is nearly starved

School of Computing Degree of Unfairness ABC Slightly better fairness, but still quite unfair

School of Computing How to Improve Fairness? Reduce inter-packet time: More TX opportunity for itself but less opportunity for others Time Adjust the inter-packet time!

School of Computing How to Improve Fairness? Increase inter-packet time: Less TX opportunity for itself but more opportunity for others Time Adjust the inter-packet time!

School of Computing Design Decisions Decision 1: Slow down offender or Speed up victim? Slow down offender!

School of Computing Design Decisions Decision 2: Increase CWmin or AIFS? CWmin!

School of Computing Impact of CWmin ABC Orange: throughput of CB Blue: throughput of AB

School of Computing Impact of CWmin ABC Large CWmin for both: Good for fairness, but bad for efficiency Default CWmin: Good for efficiency, but bad for fairness

School of Computing Impact of CWmin ABC Large CWmin for both: Good for fairness, but bad for efficiency Default CWmin: Good for efficiency, but bad for fairness

School of Computing Impact of CWmin ABC Need a dynamic CWmin adjustment algo that improves fairness without affecting much efficiency

School of Computing Key Observation ABC AB dominates CB dominates

School of Computing Key Observation ABC Good fairness Worse efficiency

School of Computing Design of FairMesh Detect the existence of unfairness Appropriately slow down the offenders by adjusting their CWmin Every node runs FairMesh protocol Every node knows local 2-hop topology

School of Computing Unfairness Detection Estimate throughput Prompt? Accurate? Use sliding window (1 s) Use per-neighbor seq. no.

School of Computing Unfairness Detection Identify victim link & offending link Fairness notion? Victim link? max-min fairness The link with the lowest thruput Offending link? The link with the highest thruput that has “conflict” with victim

School of Computing Slow Down the Offender Who Shall Take Action? A set of rules to elect “coordinators” DistributedUnique A B D C Instruct the offender to slow down

School of Computing Slow Down the Offender CWmin adjustment algorithm Propose water-discharging algo Basic idea: Gradually slow down offender until the minimum thruput cannot be improved

School of Computing Slow Down the Offender ABC An example for 2-link case D 44

School of Computing 45 Slow Down the Offender ABC An example for 2-link case D 4 MIN Offender

School of Computing 4545 Slow Down the Offender ABC An example for 2-link case D MIN better Offender

School of Computing Slow Down the Offender ABC An example for 2-link case D MIN worse Revert

School of Computing Slow Down the Offender ABC An example for 2-link case D

School of Computing Evaluation of FairMesh Testbed implementation (No BEB) ns-2 simulation (for BEB) Use saturating UDP traffic

School of Computing Evaluation of FairMesh Outline FairMesh vs & prior work BEB vs. no-BEB Lossy links & Proportional fairness Higher data rate Large-scale networks Multi-hop TCP flows

School of Computing FairMesh vs & Prior Work ABCD With BEB

School of Computing FairMesh vs & Prior Work ABCD With BEB

School of Computing FairMesh vs & Prior Work ABC With BEB

School of Computing FairMesh vs & Prior Work ABC With BEB

School of Computing FairMesh vs & Prior Work ABCD With BEB

School of Computing FairMesh vs & Prior Work ABCD With BEB

School of Computing FairMesh vs & Prior Work In summary: FairMesh achieves good fairness for the three canonical scenarios FairMesh does not negatively affect the total throughput (efficiency) Prior work are effective in some scenarios, but not all

School of Computing proportional fairness: instead of MIN, try to improve Lossy Links & Proportional Fairness ABC 42% loss max-min fairness: achieve equality, but small total thruput

School of Computing Lossy Links & Proportional Fairness ABC 42% loss max-min fairness: achieve equality, but small total thruput proportional fairness: instead of MIN, try to improve No equality, but better total thruput

School of Computing Lossy Links & Proportional Fairness In summary: FairMesh can be easily modified to support proportional fairness Users can choose the fairness notion as they want

School of Computing Large-scale Networks Arbitrarily select 20 links, from 20-node testbed

School of Computing Large-scale Networks Arbitrarily select 20 links, from 20-node testbed

School of Computing Large-scale Networks Arbitrarily select 20 links, from 20-node testbed FairMesh

School of Computing Large-scale Networks Arbitrarily select 50 links, from 50-node simulation

School of Computing Large-scale Networks Arbitrarily select 50 links, from 50-node simulation

School of Computing Large-scale Networks Arbitrarily select 50 links, from 50-node simulation FairMesh

School of Computing Large-scale Networks In summary: FairMesh can be applied for large-scale networks FairMesh achieves near-optimal max-min throughput allocation

School of Computing Evaluation Summary of FairMesh FairMesh achieves better fairness for the three canonical scenarios, as compared with and prior work FairMesh can be easily modified to support proportional fairness FairMesh can be applied for large-scale networks and achieves near-optimal max-min throughput allocation

School of Computing Thesis Outline Unfairness due to capture Pitfall of Message in Message Mechanism MAC ACK interference FairMesh Adaptive MIM MinPACK In thesis

School of Computing Dense AP Deployment

School of Computing AP Density Measurement War-walking

School of Computing Each 1-sec duration is considered as a “sample” War-walking Low speed: 1 m/s Identify an AP based on BSSID in Beacon WiFi sniffer

School of Computing War-walking Commercial area University campus Residential area

School of Computing AP Density Results Scenarios Median number of APs Channel 1Channel 6Channel 11Others Commercial 669< 1 University 865< 1 Residential 91510< 4

School of Computing Interference Mitigation Current approaches: Regulate the tx power of the MAC Data frames from AP Our key observation: MAC Acknowledgment frames from clients could also cause serious interference to neighbor cells

School of Computing Interference Mitigation Power control of Data frames cannot be applied for ACK frames, because: Data frames have tx status feedback, but ACK frames don’t have Also, ACK frames are sent at low data rate ACK frames are small

School of Computing MAC ACK Interference DATA ACK MAC ACK frames effectively extend the interference range of a hotspot AP 1 C1C1 AP 2 C2C2

School of Computing Measure the Impact of ACK Interference AP 1 AP 2 C1C1 C2C2 Campus WLAN - Cisco AP (1140 series) Clients with Atheros adapters a and n DATA Experiment Setup

School of Computing Impact of MAC ACK Interference 11n vs. 11n, UDP AP 1 AP 2 C1C1 C2C2

School of Computing Impact of MAC ACK Interference 11n vs. 11n, UDP AP 1 AP 2 C1C1 C2C2

School of Computing 11n vs. 11n, UDP Impact of MAC ACK Interference AP 1 AP 2 C1C1 C2C2

School of Computing 11n vs. 11n, UDP Impact of MAC ACK Interference AP 1 AP 2 C1C1 C2C2 Better throughput

School of Computing 11n vs. 11n, TCP Impact of MAC ACK Interference AP 1 AP 2 C1C1 C2C2 Better fairness

School of Computing 11a vs. 11n, UDP 11a 11n Impact of MAC ACK Interference AP 1 AP 2 C1C1 C2C2

School of Computing 11a vs. 11n, UDP 11n 11a Impact of MAC ACK Interference AP 1 AP 2 C1C1 C2C2 No n starvation

School of Computing Power Control of ACK DATA Sender ACK Sender Default ACK power ACK power reduced too much! Sender has to retx!

School of Computing Key idea Gradually reduce the power of ACK, until the point just before the success rate of ACK starts decreasing. Called Minimum Power for ACK (MinPACK) Challenge How can the ACK sender accurately estimate the success rate of ACK? Power Control of ACK

School of Computing Estimation of ACK Success Rate DATA Sender ACK Sender Feedback-based method 1 ACK tx 1 ACK rx 2 ACK tx 3 ACK tx 2 ACK rx Inform ACK sender 2/3 = 67% Accurate, but need to modify DATA sender!

School of Computing DATA Sender ACK Sender Passive estimation method ACK tx 2 ACK tx 2 3 ACK tx 2/3 = 67% Not perfect due to retx limit, but good enough in practice 3 Prev ACK success Prev ACK fail Prev ACK success Estimation of ACK Success Rate

School of Computing Passive Estimation for Block ACK Problem: DATA sender could send any frame that has not been acknowledged Solution: ACK sender maintains a history of frames received

School of Computing MinPACK Protocol 200 ms Time ACK tx power Initially at max power Get the max ACK success rate Φ max max Reduce if Φ > Φ max - δ Periodically set to max power to get new Φ max Revert to previous level otherwise Repeat power adjustment

School of Computing Evaluation of MinPACK Outline Gain of MinPACK - 11a vs. 11a in 20-node testbed - 11n vs. 11n in campus WLAN - 11a vs. 11n in campus WLAN Interaction with DATA power control Adaptation to client mobility

School of Computing Gain of MinPACK 20-node outdoor a testbed Arbitrarily select 38 pairs of competing links, with UDP traffic

School of Computing Throughput Gain Equal 20% gain50% gain

School of Computing Throughput Gain MinPACK does no harm Median gain is 31%

School of Computing Throughput Gain MinPACK does no harm Median gain is 31% Passive method achieves similar performance to Feedback method

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power vs. MinPACK

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power vs. MinPACK

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power vs. MinPACK

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power vs. MinPACK

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power vs. MinPACK

School of Computing Power Control of Data Frames is not Sufficient AP 1 AP 2 C1C1 C2C2 Default ACK power vs. MinPACK

School of Computing Mobility AP 1 C1C1 AP 2 C2C2 Default ACK power High throughput for both C 1 and C 2 Low throughput for C 2 Low throughput for C 1 60 m 40 m

School of Computing Mobility AP 1 C1C1 AP 2 C2C2 60 m 40 m High throughput for both C 1 and C 2 Better fairness, slightly higher total throughput Default ACK power MinPACK

School of Computing Evaluation Summary of MinPACK MinPACK is able to improve the throughput of 2 competing flows that suffer from ACK interference MinPACK is complementary to existing Data power control protocol MinPACK is adaptive to node mobility

School of Computing Thesis Outline Unfairness due to capture Pitfall of Message in Message Mechanism MAC ACK interference FairMesh Adaptive MIM MinPACK In thesis

School of Computing Thesis Contribution Summary FairMesh - Accurate assessment method for unfairness - Distributed election of coordinators - Water-discharging algorithm for max-min fairness MinPACK - Two simple and accurate methods for ACK success rate estimation - Distributed ACK power control protocol

School of Computing Thesis Conclusion Studied link-layer performance of dense WiFi networks with heavy traffic Investigated two important problems that received little attention previously - Physical-layer capture effect - MAC ACK interference Proposed effective solutions that can be immediately applied in practical deployment - FairMesh - MinPACK

School of Computing Open Issues & Future Work Open issues: - FairMesh with routing protocol and traffic priority? - FairMesh with mobility? - MinPACK with mixed downlink/uplink traffic? Future work: - Early discard of unwanted MAC frames - Aerial WiFi networks using UAV

Thank You WISH YOU A HAPPY AND PROSPEROUS NEW YEAR!

Backup Slides

School of Computing WiFi Testbed 20 nodes 3D topology

School of Computing WiFi Testbed Alix board abg cards 500 MHz CPU

School of Computing Improvement of Fairness Default ACK power MinPACK MinPACK achieves better fairness for this link pair

School of Computing MinPACK achieves better efficiency for this link pair Improvement of Fairness

School of Computing Fairness is improved for most link pairs. Some link pairs have fairness and efficiency both improved. Improvement of Fairness