Power-Aware Operator placement and broadcasting of continuous query results
description
Transcript of Power-Aware Operator placement and broadcasting of continuous query results
POWER-AWARE OPERATOR PLACEMENT AND BROADCASTING OF CONTINUOUS QUERY RESULTS
Panickos Neophytou, Mohamed Sharaf,Panos Chrysanthis, Alexandros Labrinidis
MobiDE 2010 – June 6, 2010
ΕΝΕΡΓΕΙΑΚΑ-ΕΠΙΚΕΡΔΗΣ ΤΟΠΟΘΕΤΗΣΗ ΤΕΛΕΣΤΩΝ ΚΑΙ ΕΚΠΟΜΠΗ ΑΠΟΤΕΛΕΣΜΑΤΩΝ ΕΡΩΤΗΜΑΤΩΝ ΔΙΑΡΚΕΙΑΣ
Motivation
Energy Constraints
Streams: Collection, Processing, Delivery
Social Media Events
Environment Readings
Stock Market
News Events
DSMSBroadcast
Q1Q2Q3Q4
Continuous Queries
(CQs)
Q1Q2Q3
Q1Q2Q3
Problem Definition
Q1
Q2
Q3
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
Goal:Design operator placement algorithms that balance the
tradeoff between the overall Tuning and Processing energy
at the clients.
Roadmap
Motivation/Introduction System Model
Stream Processing Model Broadcast Access Model
Operator Placement Algorithms Experiments Conclusion
Stream Processing Model
Client Tuning Energy:
Client Processing Energy:
SelectivityProjectivity
Cost in cycles
Tuning PowerProcessing PowerProcessor Speed
Streams Broadcast Model
Q1Q2Q3Q4
A broadcast is broken into cycles
Cycle:
Q1
Q3
Q4
Broadcast Organization
Streams Broadcast Model
Q1Q2Q3Q4
A broadcast is broken into cycles
Cycle:
Q1
Q4
Broadcast Organization
Q1 Q4 Q3
Q2
Streams Broadcast Model
Q1(2)
Q2(3)
Q5(1)
Q3(4)
Q4(5)
Q3
SortedBy size
Tuning Energy
Roadmap
Motivation/Introduction System Model
Stream Processing Model Broadcast Access Model
Operator Placement Algorithms Experiments Conclusion
Algorithm - MinDataCutQuery Plan:
MinDataCut gives us the minimal Broadcast Size
Minimal Edge
Tuning EnergyProcessing Energy
Clients’ Overall Energy Consumption:
Algorithm - MinPowerCutQuery Plan:
Tuning EnergyProcessing Energy
Minimal EdgeClients’ Overall Energy Consumption:
Drawbacks of MinDataCut and MinPowerCut
Q1(1)
Q4(5)
Q3(6)
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
• Oblivious to Processing costs• High processing energy
Q1(3)
Q4(5)
Q3(6)
14
• Processing-energy aware• High impact on tuning energy
MinDataCut MinPowerCut
MinPowerCut is oblivious to Broadcast Organization
BOSe: Broadcast Aware Operator Selection
Tuning EnergyProcessing Energy
Tuning EnergyProcessing Energy
Calculate the impact on:1.processing
energy2. global tuning
Query Plan (MinDataCut):
Query Plan (1 step further):
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.
BOSe: Cost-Benefit
Q1A(2)
Q2(3)
Q5(1)
Q3(4)
Q4(5)
Segment from Q1 (at Client N1)
Q1A(2)
Q1B(4.5)
Broadcast Organization (Sorted by size):
tr0 tr1 tr2
Tuning EnergyProcessing Energy
N1N2N3N4
N1 N1N4
Tuning EnergyProcessing Energy
CostBenefit
Roadmap
Motivation/Introduction System Model
Stream Processing Model Broadcast Access Model
Operator Placement Algorithms Experiments Conclusion
Experimental SetupParameter ValuesNumber of queries 20-300 (default 50)Levels per query 2-20 (default 10)Sources tuple rate 500-1000 tuples/secSources tuple size 2000-4000 bytesSelectivity 0.2-1.8, uniformProjectivity 0.5-1.5, uniformOperator costs 100*106-200*106 cycles,
Zipf
Query Workload:
Broadcast:Bandwidth 125000 bytes/sec
Mobile Clients:CPU Speed 1*109 cycles/secProcessing to Tuning power ratio
0.16
Processing to Tuning Power Ratio
BOSe always performs best
22% improvement
Scalability: Number of Queries
Scalability: Number of Operators per Query
Indexed Broadcast Model
Q1(2)
Q2(3)
Q5(1)
Q3(4)
Q4(5)
Q3
Indexed
Tuning Energy
Ix(0.5)
Processing vs. Tuning Power
53% improvement
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
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