outsourcing coordination and management of home wireless access points through an open api ashish...
TRANSCRIPT
Outsourcing Coordination and Management ofHome Wireless Access Points through an Open API
Ashish Patro
Prof. Suman BanerjeeUniversity of Wisconsin Madison
{patro, suman}@cs.wisc.edu
Outline
• Introduction• COAP Framework• Cooperation Across APs• Learning from prior
Dense residential WLANs today…Apartment Building
Apt 202Apt 201
Access Points WiFi Clients Non-WiFi devices
Dense residential WLANs today…Apartment Building
Apt 202Apt 201
Static Config
High Interference
Non-WiFi
Main observations
• Inefficient spectrum usage due to static configurations– Most APs use a single channel
Main observations
• High WiFi Interference– Average airtime utilization at the neighboring APs
increased upto 70% due to low PHY transmitters– Some links experience hidden terminal
interference from nearby APs
Main observations
• Non-WiFi interference– Non-WiFi devices do not backoff– Result in packet losses due to overlapping
Our Goal: A management framework
Determine the wireless context at its neighboring APs and WiFi channels
Determine the best remedial measure
Our Goal: A management framework
• A vendor-neutral API• A centralized framework • A cloud-based management service• Using Software-Defined approach
How to manage different residential wireless APs from different vendors?
Outline
• Introduction• COAP Framework• Cooperation Across APs• Learning from prior
Coordination framework for Open APs
Apartment Building
AP
AP
COAP Controller
ISP x
ISP yInternet
Last Hop ISPs
Cordless Phone Laptop
Wireless TV
Smartphone
LaptopMicrowave
Oven
Measure
Measure
API
API
Config
Config
COAP framework
COAP framework
Implemented OpenFlow modules
APConfigManager: Receive configuration commands from the controller
DiagnosticStatsReporter: Report detailed wireless statistics to the controller
BasicStatsReporter: Report aggregate wireless statistics to the controller
COAP controller modules
Access Points Controller Wireless OpenFlow
COAP framework implementation
Access Points Controller Wireless OpenFlow
COAP framework implementation
Airshark (IMC 2011)
Packet capturing & parse the packet headers to obtain link level statistics
Non-WiFi device detection capability using commodity WiFi cards
BasicStatsReporter
&DiagnosticStats
Reporter
COAP framework implementation
Access Points Controller Wireless OpenFlow
COAP framework implementation
Access Points Controller Wireless OpenFlow
COAP framework implementation
• Transmit wireless configuration updates from the controller to the APs– switch channel– throttle airtime
AP
Controller
COAP deployment
• 12 OpenWrt based COAP APs– Used as private APs– Use a secondary NIC on the APs to collect airtime
utilization information across all channels in a round robin fashion.
• 30 WiSe APs
WiSe deployment (30 APs)
Building 1: APs 1 – 14Individual Access Point
per apartment
Building 2: APs 25 – 30Deployment in common
areas
Others: APs 15 – 24Across different
homes
Ran deployment over 8 months
Outline
• Introduction• COAP Framework• Cooperation Across APs• Learning from prior
Cooperation across APs - Channel
Controller
Configuration
Administrator
Cooperation across APs - Channel
COAP Controller
Measure MeasureConfiguration
Full view of the spectrum…Can the controller leverage spatio-temporal locality of
nearby APs for better channel selection?
feasibility
feasibility
COAP Controller
Full view of the spectrum…Can the controller leverage spatio-temporal locality of
nearby APs for better channel selection?
CDF of the Pearson’s correlation coefficient for time-seriesper-channel airtime utilization observed by neighboring AP pairs
more than 60% of nearby AP pairs (RSSIs > -55 dBm) exhibited a high correlation coefficient
Performance improvements
It shows that the dynamic "airtime-ware“ scheme performed better than a random channel assignment scheme for 10 out of the 12 APs
Cooperation across APs - Airtime
SetAirtimeAccess( transmit_bitmap, slot_duration)
Channel congestion caused by nearby AP traffic
Hidden terminal style interference
API
Problems
To mitigate these scenarios by controlling the airtime access patterns of the interfering APs
Airtime management - API
Controller divides time
into small "slots"
Enable/disable transmissions of COAP APs on a per-slot
basis
API
Limit a COAP AP’s access to certain slots
Avoid overlapping
Throttle(APx)
Slot( Apx, APy)
Airtime management - API
Testbed evaluation
802.11n based COAP APs
clients
×𝟔
×𝟔
hidden terminal client-side interference
channel congestion experienced by APs
Two scenarios
Testbed evaluation
802.11n based COAP APs
clients
3 links consisted of HTTP based video traffic
3 links using iperf traffic
6 links
TCP throughput Metrics?MAC loss rates Frame drop rate
Hidden terminal scenario
×𝟑10 Mbps HD video
×𝟑10 Mbps traffic
802.11 DCF Slot(APx,APy)VS
Hidden terminal scenario
DCF scenario: all three video flows experienced high MAC layer losses
Slot scenario: throughput improved of all video links
Mitigating channel congestion
10 Mbps 5 Mbps5 Mbps
DCF scenario: the high bitrate video link experienced high frame drop rates
3 links consisted of HTTP based video traffic
3 links using iperf
traffic
Mitigating channel congestion
10 Mbps 5 Mbps5 Mbps
the performance of high bitrate video link improved due to the higher throughput achieved by the link
3 links consisted of HTTP based video traffic
3 links using iperf
traffic
throttled to 50%
Outline
• Introduction• COAP Framework• Cooperation Across APs• Learning from prior
Learning to predict
COAP Controller
Prior wireless activity
logs
learning "context-related"
information
Predicting future non-WiFi activity
Predict traffic characteristics
Modeling non-WiFi activity
Airshark
activity vector
Each element (ci) in the vector records the average number of daily instances of non-WiFi activity observed during a time "bin period"
Modeling non-WiFi activity
"Activity vectors" for microwave oven activity observedby three different COAP APs (2 weeks).
Predicting non-WiFi activity
• For a given time span of k days, using per-AP activity data (total of d days), we obtained a sequence of activity vectors ( e.g., k = 30)
• Computed the per-AP Pearson’s correlation coefficient between consecutive activity vectors and averaged them
Predicting non-WiFi activity
• For a given time span of k days, using per-AP activity data (total of d days), we obtained a sequence of activity vectors ( e.g., k = 30)
• Computed the per-AP Pearson’s correlation coefficient between consecutive activity vectors and averaged them
Used these sequences of activity vectors to determine the predictability of future non-WiFi activity based on
the most recent record of non-WiFi activity
Learning client and traffic context
• Bias in traffic usage profile by device type• Impact of device usage characteristics• Platform specific traffic behavior
Client and traffic context information can be helpful
Learning client and traffic context
Burst properties
Session properties
consecutive active periods with a gap less than 10
seconds
a sequence of consecutive traffic bursts with a gap of
less than 5 minutes
duration, downloaded bytes, the average and maximum download speed
gaps, duration, bytes downloaded and download speeds
Predicting traffic characteristics
COAP AP context a collection of the following traffic and device related features
AP ID, client device id, trafficsource id, time of day, day of week
Machine learning tool, Weka …
predict the burst and session related properties
compared with non-context predict
Predicting traffic characteristics
Q & AThank you!