KAI ST Distributed cross-layer scheduling for Distributed cross-layer scheduling for In-network sensor query processing In-network sensor query processing PERCOM 2006 2006. 11. 23(THU) Lee Cheol-Ki Network & Security Lab.
KAIST
Distributed cross-layer scheduling for Distributed cross-layer scheduling for
In-network sensor query processingIn-network sensor query processing
PERCOM 2006
2006. 11. 23(THU)
Lee Cheol-Ki
Network & Security Lab.
22 / 17 / 17 컴퓨터 구조 특강
ContentsContents
Introduction
Related work
System overview
Schedule construction
Schedule execution
Evaluation
Conclusion and future work
33 / 17 / 17 컴퓨터 구조 특강
Introduction(1/2)Introduction(1/2)
In-network sensor query processing systemsData acquisitional applications of wireless sensor networks
Inject queries into the network
The networked sensor nodes work together to process the queries
Send back the query results
Critical issue : To reduce power consumption
Major sources of energy waste
Idle listening, over hearing, collision and control packet overhead
Immediate solution
To make the nodes sleep as much as possible
Focus on tree networks (similar to the existing schemes)
44 / 17 / 17 컴퓨터 구조 특강
Introduction(2/2)Introduction(2/2)
Process of systemsA node first checks Applicable Transmission Slots
Sends it to parent
The parent sends the assigned transmission timing
The node arranges it about tasks in all layers
Execution the schedules
55 / 17 / 17 컴퓨터 구조 특강
Related work(1/2)Related work(1/2)
Scheduling has been applied to different layers
TinyDB in query layer
S-MAC in MAC layer
Centralized approach (Florens)
Sink allocate transmission timing to all other nodes in WSN
Sink know the network topology → difficult, high cost
Destributed approach
Flexible Power Scheduling (FPS)
Distributed on-demand power-management protocol for tree networks
Reducing the collision between siblings
But, not reducing the collision between neighbors that are not siblings
66 / 17 / 17 컴퓨터 구조 특강
Related work(2/2)Related work(2/2)
Sichitiu
The source node broadcast a special route setup packet with neighbors
Temporary schedule → permanent schedule (after arriving at sink node)
Frequent collision → dead nodes → inaccurate query result
In this paperConsidering all tasks query injecting, computation, aggregation
Consecutive sleeping and transmission timing to nodes
All layers of a query processing system are involved
77 / 17 / 17 컴퓨터 구조 특강
System overviewSystem overview
In-network sensor queryprocessing system
Construct routing tree
Sink inject a query
Nodes construct schedules
Sink broadcasts synch-signalBottom-up report to sink
Run following the schedules
Scheduling moduleSchedule construction
Time synchronization
Schedule execution<Architecture of scheduling module>
<A typical WSN setup>
88 / 17 / 17 컴퓨터 구조 특강
Schedule construction(1/4)Schedule construction(1/4)
Schedule construction module
Constraints for schedule construction
<Example of constraint >ⅲ
99 / 17 / 17 컴퓨터 구조 특강
Schedule construction(2/4)Schedule construction(2/4)
ConstraintsFocus on collisions at the parent node of a sender
Complete collision-free scheduling → high overhead, impractical
In practical
D and E sends packet simultaneously Collision occurs in B⇒
But, don’t affect to the query results
<An example of the collision>
1010 / 17 / 17 컴퓨터 구조 특강
Schedule construction(3/4)Schedule construction(3/4)
Notation
Variables
CTS (Children Transmission Slots), ATS (Applicable Transmission Slots),TSI (Transmission Slot Information), CATS
Type of time slots
PL/R(processing-listening/receiving), Q/M(Query injection/maintenance),Sleeping slots, transmission slots
Procedures
A node, to determine ATS (DetermineApplicableTransmissionSlots)
Sends the packet to the parent
Parent node, CTS for children (AllocateChildrenTransmissionSlots)
A node receive TSI from the parent node
Sends ACK
1111 / 17 / 17 컴퓨터 구조 특강
Schedule construction(4/4)Schedule construction(4/4)
Complete scheduleConstructSchedule arrangesthe time slots for PL/R, Q/M
<Example of complete schedules>
1212 / 17 / 17 컴퓨터 구조 특강
Schedule executionSchedule execution
Execution involves the timing control of all layersMAC layer checks whether it is time for transmission
Earlier than allocated transmission time
Copy the messages to a memory buffer and wait
Timer setup (automatically transmission)
Waiting time should be shorter than the interval of 2 transmissions
Routing layer
Control the timing of transmission for the route maintenance messages
1313 / 17 / 17 컴퓨터 구조 특강
Evaluation(1/4)Evaluation(1/4)
Scheme comparison (DCS, FPS, SS)The experimental setup
The same experimental setup as that Sichitiu used
100 nodes in an 80m*80m rectangular area, Trans-range : 25m
Query processing performanceThe experimental setup
10 Crossbow MICA2 motes, optimized TinyDB, original TinyDB
HP-4156 oscilloscope for measuring the power consumption
VMNet simulator
1414 / 17 / 17 컴퓨터 구조 특강
Evaluation(2/4)Evaluation(2/4)
Scheme comparisonScheduling overhead
Time of constructing a schedule
Schedule comparison
Number of dead node
Non-scheduled node, conflict-scheduled node
AFS (Average frequency of switching between an active slot and a sleeping slot)
<Scheduling overhead>
<Number of dead nodes> <Average frequency of switching>
1515 / 17 / 17 컴퓨터 구조 특강
Evaluation(3/4)Evaluation(3/4)
Query processing performanceSchedule for in-network aggregation
VMN : Emulated network in VMNet
Query 1
SELECT avg(light) FROM sensors
Sample intervals
2, 10, 60 seconds
Comparison optimized TinyDB with original TinyDB
42%, 67%, 75% at sample interval of 2s, 10s and 60s
<The VMN topology with schedules>
1616 / 17 / 17 컴퓨터 구조 특강
Evaluation(4/4)Evaluation(4/4)
Power consumption improvement
Difference between measurement in VMNet and that in the real WSN within ±15%
Difference in the topologies
Measurement errors of the oscilloscope
<Power consumption improvement>
1717 / 17 / 17 컴퓨터 구조 특강
Conclusion and future workConclusion and future work
Propose scheduling scheme for Power efficiencyReduce the number of dead nodes
Reduce Switching frequency
In-efficient to reconstruct a schedule whenA change in the network topology
A new query arrives