SwissQM Nextgeneration Data Processing in Sensor Networks

1
SwissQMNext-generation DataProcessing in Sensor
Networks

• YANG Jun
• 2008.4.7

Some of the slides are from the authors René
Müller, Gustavo Alonso, and Donald
Kossmann Department of Computer Science, ETH
Zurich
2
Outline

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

3
Introduction(1)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Wireless Sensor Networks
• Collection of Sensor Nodes
• Sensor Node
• A Processor
• A Radio Unit
• A bunch of Sensors
• A Battery
• Ad Hoc Networks

4
Introduction(2)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Sensor Network Applications

Structural monitoring
agriculture
Urban warfare
5
Introduction(3)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Typical Application Scenario
• Distribute a number of sensor nodes that measure
  property X.
• Add a PC/PDA/ that will record the measured
  data.
• Connect a the PC/PDA/ with one sensor.
• Nodes build an Ad-hoc Network (e.g., routing
  tree).
• Let them send (periodically) along tree to the
  PC/PDA/

6
Motivation(1)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Limited Resource of Sensor Nodes
• Flash-Memory, RAM, Radio, Battery
• Programming on Sensor Nodes for Data Processing
• In practice very low-level
• The dirty way NesC, TinyOS, etc.
• The refined way Queries (TinyDB/Cougar like)
• Very platform/target specific
• Cumbersome
• Professional Debugging
• Only 3 LEDs for status information of the program
• Expensive debugging kits (e.g. JTAG) lead to
  inner workings of processor

7
Motivation(2)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• We needed something better!
• Something Like a System
• Language independent (SQL, XQuery, Java, new
  languages, Webservices )
• Turing complete (can compute every
  Turing-computable function)
• User-defined functions
• Capable of pushing down complex processing
  functions all the way to the sensors

8
Motivation(3)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Solution
• Virtual Machine tailored to data acquisition in
  sensor networks
• Virtual Machine Example JVM

9
SwissQM

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Scalable WIreleS Sensor Query Machine

10
SwissQM Overview

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Design Considerations
• Sensor Network Gateway
• SwissQM On The Sensor Node
• SwissQM Programs
• Bytecode Instructions
• Multi-programming
• Messaging In The Network
• In-network Aggregation
• Program Propagation

11
SwissQM(1)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Design Considerations
• Separation of sensors and external interface
• Language-independent
• No application specific functionality in sensors
• Dynamic, multi-user, multi-programming
  environment
• No imposed restrictions on the query submission
  and the number of queries running concurrently
• Optimized use of the sensors
• The sensors only process capturing, aggregating,
  forwarding data.
• Extensibility
• Allow user defined functions
• Possible to extend itself dynamically as needed

12
SwissQM(2)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Sensor Network Gateway
• Be exploited as much as possible
• Everything that can be done at the gateway is
  done
• Provide all external interfaces query
  optimisation and compiltion facilities

13
SwissQM(3)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Outlook of SwissQM Gateway System

14
SwissQM(4)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• SwissQM is a Virtual Machine on sensor node
• Virtual Machine
• Instruction Set
• Elementary Instruction Application-specific
  extension
• Functionality can be implemented
• In bytecode
• As additional instruction

15
SwissQM(5)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• SwissQM on sensor node

• Stack-based VM
• Integer Arithmetic ONLY
• 59 bytecode instructions
• 40 VM core
• 19 sensor network specific
• Transmission Buffer
• Buffer the data received and the data to be sent
• Synopsis
• Store the aggregation state
• Merge the data from its children and the data
  itself
• Sensor Interface

16
SwissQM(6)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• SwissQM Programs
• Written in SwissQM bytecode (Java-like)
• Streaming data

• Consists of 3 sections
• Init section program initialization
• Delivery section invoked periodically
• Reception section called when data received from
  children

17
SwissQM(7)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Bytecode Instructions
• VM Instructions
• 16 Stack Instructions
• 11 Arithmetic and Logical Instructions
• 13 Control Instructions
• SwissQM Core Instructions
• 9 Buffer Instructions
• 7 Sensor Instruction
• 2 Transmission Instructions
• Application-specific
• 1 In-network Aggregation

18
SwissQM(8)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

19
SwissQM(9)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Multi-programming
• Support on two layers
• Merge different user queries into a virtual query
• Multi-programming in the QM
• In one sampling interval, sequentially execute
  programs
• Sample periods in the order of seconds
• Data transmission in the order of
  milliseconds
• Program execution in the order of microseconds
• Trade-off
• Number of programs VS Memory available for each
  program
• Can run over hundred user queries

20
SwissQM(10)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Messaging in the Network
• Message Types
• Program Messages
• Result Messages
• Command Messages (Node Reset, Program Stop,)
• Traffic Pattern
• Root to all Program and Command Messages
• All to Root Result Messages
• Local Broadcasts Routing updates and time
  synchronisation

21
SwissQM(11)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• In-network Aggregation
• In-network aggregation can reduce the number of
  messages (save energy)
• TAG Approach Tiny AGgregation Madden OSDI02
• Example What is the average temperature in the
  network?

22
SwissQM(12)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• State of AVERAGE aggregation
• Sum
• Count
• In-network aggregation
• Tree topology
• Only one message per link
• Children send data before parent
  (Scheduling/Timesync)

23
SwissQM(13)

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• Program Propagtion
• Restricted message size in ad-hoc sensor
  networks.
• TinyDB uses 49 bytes message size but large data
  losses are observed in its deployments
• SwissQM opts for the smaller 36 bytes message
  size and fragment messages are supported.
• Use the same protocol as TinyDB in default.

24
Application Examples

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

Example 1 Get the nodeid of the node whose
temperature exceeds 60?

• QM program
• .section delivery, _at_10s
• get_temp read temperature sensor
• ipushb 60
• if_icmple end skip if templt60
• get_nodeid read nodes ID
• istore 0 store it at pos. 0
• send_tb send transmission buffer
• end
• .section reception
• send_tb forward tuple from child

25
Application Examples

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

Example 2 Assume that Light Sensor only produces
unprocessed raw values from the A/D converter and
light sensor has an offset reading that needs to
be accounted for. Goal Average temperature
readings of sensor nodes that have similar
brightness readings.
26
Conclusions and Future Work

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• SwissQM
• Flexible programming platform for data
  acquisition tasks in sensors networks
• Offers a well defined and extensible bytecode
  language
• Can be used as compilation target for higher
  programming and query languages (XQuery, SQL)
• Increases abstraction level at the network
  interface
• Powerful instruction set -gt short programs -gt
  eases dissemination
• Future Work
• Future Sensor
• May have more memory and CPU power
• But radio bandwidth and reliability still an
  issue
• Cost-efficiency
• XML and XQuery support

27
As a Reviewer

• Background Introduction
• Motivation
• A Solution SwissQM
• Application Examples
• Conclusions and Future Work
• As a Reviewer

• SwissQM v.s TinyDB
• Borrows many ideas from TinyDB, but improved a
  lot due to the design

• Seems too little evaluations on the performance
• Itll be more convincing If they did a long-term
  experiment to evaluate the performance of running
  TinyDB on top of SwissQM compared with running
  TinyDB itself

28
End

• QA
• Thank you