This document presents three algorithms for power-aware operator placement and broadcasting of continuous query results in a distributed stream processing system. The first two algorithms, MinDataCut and MinPowerCut, aim to minimize data transfer or processing power consumption but are not optimal. The third algorithm, BOSe, considers both tuning and processing energy costs to select operator placements. Experiments show BOSe improves energy efficiency by 53% over centralized processing. Future work includes supporting operator and query sharing to further optimize the energy-response time tradeoff.

1. POWER-AWARE OPERATOR
PLACEMENT AND
BROADCASTING OF
CONTINUOUS QUERY RESULTS
Panickos Neophytou, Mohamed Sharaf,
PanosChrysanthis, AlexandrosLabrinidis
MobiDE 2010 – June 6, 2010

2. Motivation
Energy
Constraints

3. Streams:
Collection, Processing, Delivery
Social Media
Events
Environment
Readings DSMS Broadcast
Stock
Market
Q1
Q2
News Continuous
Events Q3 Queries
(CQs)
Q4

4. Problem Definition
Goal:
Designoperator placement algorithms
that balance the tradeoff between the
overall Tuning and Processing energy at
Tuning Energy
the clients. Processing Energy
Q1
Q1 Tuning Energy
Q1 Processing Energy
Q2 Q2
Q2
Q3
Q3
Q3 Tuning Energy
Processing Energy

5. Roadmap
 Motivation/Introduction
 System Model
 Stream Processing Model
 Broadcast Access Model
 Operator Placement Algorithms
 Experiments
 Conclusion

6. Stream Processing Model
Selectivity Tuning Power
Projectivity Processing Power
Cost in cycles Processor Speed
Client Tuning Energy:
Client Processing Energy:

7. Streams Broadcast Model
A broadcast is broken into cycles
Q1 Q1 Broadcast Organization
Q2
Q3 Q3
Q4 Q4
Cycle:

8. Streams Broadcast Model
A broadcast is broken into cycles
Q1 Q1 Broadcast Organization
Q2 Q2
Q3
Q4 Q4
Cycle:
Q1 Q4 Q3

9. Streams Broadcast Model
Q5 Q1 Q2 Q3 Q4 Sorted
(1) (2) (3) (4) (5) By size
Q3 Tuning Energy

10. Roadmap
 Motivation/Introduction
 System Model
 Stream Processing Model
 Broadcast Access Model
 Operator Placement Algorithms
 Experiments
 Conclusion

11. Algorithm -
MinDataCut
Query Plan:
Minimal Edge
Clients’ Overall Energy Consumption:
Tuning Energy
Processing Energy
MinDataCutgives us the minimal Broadcast Size

12. Algorithm -
MinPowerCut
Query Plan:
Minimal Edge
Clients’ Overall Energy Consumption:
Tuning Energy
Processing Energy

13. Drawbacks of MinDataCut and
MinPowerCut
MinDataCut MinPowerCut
Tuning Energy Tuning Energy
Processing Energy Processing Energy
• Oblivious to Processing costs • Processing-energy aware
• High processing energy • High impact on tuning energy
Q1 Q4 Q3 Q1 Q4 Q3
(1) (5) (6) (3) (5) (6)
4 1
MinPowerCut is oblivious to Broadcast Organization

14. BOSe: Broadcast Aware Operator
Selection
Query Plan (MinDataCut): Query Plan (1 step further):
Tuning Energy Tuning Energy Calculate the impact on:
Processing Energy Processing Energy 1. processing energy
2. globaltuning
1. Start from the MinDataCut point.
2. For each query, calculate the amount of energy reduction provided by
each segment of operators if it were brought back to the server.
3. Bring back the one segment with the maximum reduction.
4. Repeat until no more energy reduction is attainable.

15. BOSe: Cost-Benefit
Segment from Q1 (at Client N1) Tuning Energy
Processing Energy
Q1A Q1B Benefit Cost
(2) (4.5)
tr0 tr1 tr2 N1 N1
N4
N1
N2
N3
N4
Tuning Energy
Processing Energy
Broadcast Organization (Sorted by size):
Q5 Q1A Q2 Q3 Q4
(1) (2) (3) (4) (5)

16. Roadmap
 Motivation/Introduction
 System Model
 Stream Processing Model
 Broadcast Access Model
 Operator Placement Algorithms
 Experiments
 Conclusion

17. Experimental Setup
Query Workload:
Parameter Values
Number of queries 20-300 (default 50)
Levels per query 2-20 (default 10)
Sources tuple rate 500-1000 tuples/sec
Sources tuple size 2000-4000 bytes
Selectivity 0.2-1.8, uniform
Projectivity 0.5-1.5, uniform
Operator costs 100*106-200*106 cycles, Zipf
Broadcast:
Bandwidth 125000 bytes/sec
Mobile Clients:
CPU Speed 1*109 cycles/sec
Processing to Tuning power ratio 0.16

18. Processing to Tuning Power Ratio
22% improvement
BOSe always performs best

19. Scalability: Number of Queries

20. Scalability: Number of Operators
per Query

21. Indexed Broadcast Model
Ix Q5 Q1 Q2 Q3 Q4 Indexed
(0.5)
(1) (2) (3) (4) (5)
Q3 Tuning Energy

22. Processing vs. Tuning Power
53% improvement

23. Conclusions
 3 power-aware operator placement
algorithms for broadcasting CQ results
 BOSe algorithm improves by 53% over
centralized processing
 Future:
 Support sharing of operators
 Support sharing of queries
 Study the tradeoff between Energy and
Response Time

24. Thank you – Questions?
 Advanced Data Management
Technologies Laboratory
 http://db.cs.pitt.edu
 Part of AQSIOS project:
 NSF GRANT IIS-0534531
 NSF career award IIS-0746696