Stream Processing in Networks of Smart Devices

1
Stream Processing in Networks of Smart Devices
Holger Ziekow, Lenka Ivantysynova

• Institute of Information Systems
• Humboldt University of Berlin, Germany

2
Stream Processing on Smart Items (Motivation)

• Business Applications aim to integrate data from
  smart devices (e.g. sensors and RFID data)
• Sensor and RFID data have the properties of data
  streams (Stream, Aurora)
• Processing streams on the device layer can extend
  device lifetime by reducing communication
  (Courgar)

3
Stream Processing on Smart Items (Motivation)

• Smart devices are commonly battery powered and
  therefore very energy constrained
• Communicating is much more energy consuming than
  calculations (sending 1 bit 1000 CPU
  instructions)? Data processing in the network is
  favorable

How to map?
4
Querying in the Network

• Challenges
• Devices vary in
• Position in the network.
• Free memory.
• Operators network position influencesenergy
  consumption.
• Memory influences data accuracy.
• Mapping problem is NP hard.(Rectilinear Steiner
  Tree Problem)
• Query plans may have to be modified.

5
Stream Processing on Smart Items (Optimized
Mappings)

• Finding an optimal mapping is an NP-hard problem
• We define a metric to measure a mappings quality.
  This metric can be used in optimization algorithms

Energy
Data quality
Parameters to mutate

• Target devices for the query operator (m) for the
  given query plan (q)
• Operators in the query plan (q) which can
  subsequently be calculated

6
Stream Processing on Smart Items (Test Results)

• We used our metric and a genetic algorithm to
  find good mappings of query plans
• Tests show that good results can be found
  relatively fast
• In manual checks the generated mapping can be
  proven as reasonable

Cost
Approximation using a genetic algorithm
Optimization steps
7
Stream Processing on Smart Items (Test Results)
Query AGG(JOIN(Src1,Src2))
Memory Usage AGG 50 JOIN 70
Memory
8
Stream Processing on Smart Items (Test Results)
Query AGG(Src1,Src2,Src3,Src4)
Mapped Query AGG(AGG(Src1,Src4),AGG(Src2,Src3))
9
Future Work

• Additional optimization parameters
• Message delay.
• Value based errors.
• Node specific energy consumption.
• Tuning the optimization algorithm.
• Integration of different routing algorithms.