u niversity of m assachusetts, a mherst department of computer science multi-user data sharing...
Post on 21-Dec-2015
217 views
TRANSCRIPT
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Multi-user Data Sharing System in Radar Sensor
NetworksMing Li, Tingxin Yan, Deepak Ganesan, Eric Lyons,
Prashant Shenoy, Arun Venkataramani, and Michael Zink
Department of Computer ScienceUniversity of Massachusetts, Amherst
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Emerging Rich Sensor Networks
Richer energyTethered power
High data rateMany MB/second
Diverse users/applications needsE.g. First responders, Commuters, Insurance, for traffic monitoring
Radar Sensor Network Camera Sensor Network
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
CASA Radar Sensor NetworksDensely monitoring the lower troposphere
Tornado, storm, flood, …
High rate sensor streams300MB per radar scan every 30 seconds
Stream-based systemData processing is done on proxy
Wide-area wireless mesh network
Multiple, diverse user needsEmergency personnel, meteorologist, other…
InternetInternet
Proxy
Emergency Personnel Meteorologist
NormalUser
Tornado Detection
ReflectivityOverview Precipitation
Data S
tream
Data Processing
Query
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Challenges in Multi-user WSNs
A B
C D
A B
C D
TornadoDetection
3D Assim Wind Dir
Estimation
NWSEmergencyPersonnel
Researcher
Limited network resourcesBandwidth << Data needs
Diverse end user query needs
Diverse data quality metricsTornado: location error.Wind direction: direction error
Different spatial areas of interest
Wind direction: overlapping area between radars
Different data fidelity needsTornado detection > 3D assimilation
Different priorities and deadlinesPriority:NWS > Em. MgrDeadline: Em. Mgr < NWS
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Problem Statement
A B
C D
A B
C D
TornadoDetection
3D Assim Wind Dir
Estimation
NWSEmergencyPersonnel
Researcher
How to design next generation wireless radar sensor networks to:
Jointly optimize for different data quality metrics and different priorities and deadlines of different users
Share bandwidth and data across different users
Adapt gracefully to bandwidth dynamics
Prioritize important data during critical weather events.
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Key Ideas in Multi-User Data Sharing
Utility-driven transmission scheduling to prioritize data transmission and maximize overall utility
Progressive compression to minimize bandwidth usage and adapt to bandwidth fluctuation
Global transmission control to prioritize data transmission among radars
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Outline
MotivationKey IdeasProgressive CompressionUtility-driven Transmission SchedulingGlobal Transmission ControlEvaluationSummary
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Radar
Utility-basedScheduler
ProgressiveCompressor
…
Progressive Compressor
SPIHT algorithm [set partition in hierarchical trees]
Wavelet-based encoder, preserves important features of interest for meteorologistAdapts to bandwidth fluctuationMost important data is transmitted first
Raw Data
NWS
EmergencyPersonnel
Researcher
Err
Data Size
Decode error of SPIHT
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Progressive Compressor
Operation of Progressive compressor:
Split scan into spatial regions, each with a set of queries associated with it.
Generate a separate stream for each spatial region.
Radar
Utility-basedScheduler
ProgressiveCompressor
…Raw Data
NWS
EmergencyPersonnel
Researcher
TornadoDetection
Assimilation
Wind DirEstimation
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Utility-based Scheduler
Utility-based SchedulerDecides which packet offers greatest improvement to overall utility
Key questionsHow to determine utility of packet to a query?How to aggregate utilities across diverse queries?How to schedule packets based on their utilities?
Radar
Utility-basedScheduler
ProgressiveCompressor
…Raw Data
NWS
EmergencyPersonnel
Researcher
TornadoDetection
Assimilation
Wind DirEstimation
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Utility of a Packet to a Query
What does scheduler haveMarginal Data MSE
What does scheduler needMarginal query quality
How to map data MSE to query qualityTrain a mapping function a priori using sample data sets
Distance between detected tornado and the actual one
Intensity of detected tornadoIntensity of actual tornado
SPIHTData Stream
Error Trace
Application level
Networking level
Tor Err
Data MSE
Training
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Aggregate Utility across Diverse Queries
Aggregated utility is a weighted combination of utilities for each query
Weightquery=f(query_priority, query_deadline)
UtilAgg=sum(Weightquery*Utilquery)
Utility1
Utility2
Tornado Utility
U tor, W 1
U asm, W 2
Utility
Query1
Tornado Detect
Weight1
Assim Utility
Query2
Assimilation
Weight2
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Schedule Packets based on Utility
Schedules packet with the highest utility
Optimal if utility function as a function of data size is concave
Scheduler
P1
P2
P3
U1>U2>U3
P1
Utility
Data Size
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Outline
MotivationKey IdeasProgressive CompressionUtility-driven Transmission SchedulingGlobal Transmission ControlEvaluationSummary
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Global Transmission ControlProblem:
Radar with critical data may not get sufficient bandwidth
Solution: Proxy pauses streams that are achieving low/no utility gain
Proxy
TornadoDetect Assim
High Util Low Util
GlobalTransmission
Control
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Outline
MotivationKey IdeasProgressive CompressionUtility-driven Transmission SchedulingGlobal Transmission ControlEvaluationSummary
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Evaluation
Testbed13 MacMini as adhoc mesh nodes 3-hop topology
Data SetsReal data traces from Oklahoma radar testbedSimulated data by ARPS(Advanced Regional Prediction of Storms)
Query PatternTornado Detection, 3D assimilation and Wind Direction assimilation queries arrive in a round robin manner. Deadlines are chosen from a Poisson distribution with mean at 30 seconds.
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Determining the Utility Function
Tornado detection needs more accurate data than 3D assimilation.
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Performance of Utility-driven Scheduler
Compare utility-driven scheduler to random scheduler
2x
The utility-based scheduler achieves 2 times higher utility than the random scheduler
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Scalability Demonstrates that our system as a whole scales well with network size
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Scalability Baseline System
Averaging compression
Bandwidth estimationEstimate bandwidth for next epoch (30 secs) based on history of bandwidth from previous epochs.Conservative estimate to ensure that compressed scan can be transmitted in the 30 seconds.
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Scalability Demonstrates that our system as a whole scales well with network size
3x38%
30%
10x
4%
Our system achieves more than 10 times higher utility than the baseline system
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Related WorkMulti-query optimization in sensor network
SQL Queries and simple aggregates: Trigoni, et. al [DCOSS 2005]We have more complex data processing requirements.
Utility-driven network design in sensor networks and InternetSORA [NSDI05], Kelly et al [JORS98]Does not address application-level data quality metrics and data sharing between users
Global transmission controlConflict Graph – Jain et al [Wireless Network 05], Rate control – Rangwala et al [Sigcomm06]We use application level utility of data to control transmissions.
Radar sensor networksSchedules radar scanning strategy to satisfy end-user needs Zink et al [EESR05].
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Summary
Illustrates new challenges in next-gen radar sensor networks
Design and implementation of a multi-user data sharing system that:
Gracefully degrades utility under bandwidth fluctuations by using progressive compression Utility-driven packet scheduling based on end-user data quality metrics, priorities, and deadlines.Globally prioritizes data transmission across radars.
Results show one order of magnitude improvement in application utility over existing radar data transmission system.
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
The End
Questions?
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Future work
Joint radar sensing, bandwidth and energy optimization
Extend system to other types of WSNs like camera sensor networks.
Design a hop-by-hop bulk transfer protocol that optimizes radar data transfer
Explore rate control and bandwidth allocation for global transmission control
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Radar
MUDS Overview
Utility-basedScheduler
ProgressiveCompressor
…
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Utility-based Scheduler
Utility-based SchedulerSchedules packets between different streams based on the overall utility of each packet to the queries
Each stream is shared by multi-queries and each application has different data needs
How to determine utility of packet to an application?How to aggregate utilities across diverse queries?How to schedule packets based on their utilities?
Radar
Utility-basedScheduler
ProgressiveCompressor
…Raw Data
NWS
EmergencyPersonnel
Researcher
TornadoDetection
Assimilation
Wind DirEstimation
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Utility-based Scheduler
Utility-based SchedulerSchedules packets between different streams based on the overall utility of each packet to the queries
Each stream is shared by multi-queries and each application has different data needs
How to determine utility of packet to an application?How to aggregate utilities across diverse queries?How to schedule packets based on their utilities?
Radar
Utility-basedScheduler
ProgressiveCompressor
…Raw Data
NWS
EmergencyPersonnel
Researcher
TornadoDetection
Assimilation
Wind DirEstimation
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Adaptation to Bandwidth Fluctuation
Compare progressive SPIHT to non-progressive SPIHT
4x
Progressive SPIHT achieves up to 4 times lower data MSE than the non-progressive scheme
Bandwidth Estimator
SPIHT Data Stream
SPIHT Compressed Data
Progressive
Non-Progressive
Bandwidth TraceMoving Window
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Scalability Baseline System
Bandwidth estimation
Averaging compression
Scan 1 Scan 2 Scan 3 …
Transmit 1 Transmit 2 Transmit 3 …
Sense
Transmit
CDF
Bandwidth
5%
UUNIVERSITY OF NIVERSITY OF MMASSACHUSETTSASSACHUSETTS, A, AMHERST • MHERST • Department of Computer Science Department of Computer Science
Radar
Utility-basedScheduler
ProgressiveCompressor
…
Progressive Compressor
SPIHT algorithm [set partition in hierarchical trees]
Wavelet-based encoder, preserves important features of interest for meteorologistAdapts to bandwidth fluctuationMost important data is transmitted first
Raw Data
NWS
EmergencyPersonnel
Researcher
Err
Data Size
Decode error of SPIHT
TornadoDetection
Assimilation
Wind DirEstimation