Skyline Queries Against Mobile Lightweight Devices in MANETs Zhiyong Huang 1 Christian S. Jensen 2 Hua Lu 1 Beng Chin Ooi 1 1 2 National University of Singapore, Singapore Aalborg University, Denmark
Outline • Introduction • Problem Definition • Skyline Queries in MANETs • Optimizations on Mobile Devices • Experimental Studies • Conclusion
Introduction • Skyline query – Operator based on dominance • MANET – Self-organizing, wireless mobile ad-hoc networks – Physical environment of this work • Lightweight devices
Skyline Queries in MANETs • Assumptions – Each resource-constrained device holds a portion of the entire dataset – Devices communicate through MANET – A mobile user is only interested in data of a limited geographical area, though the query involves data stored on multiple mobile devices
Example • M 1 to M 4 hold different hotel relations • M 2 is interested in cheap and good hotels within the circle area M 1 M 2 M 3 M 4
Outline • Introduction • Problem Definition • Skyline Queries in MANETs • Optimizations on Mobile Devices • Experimental Studies • Conclusion and Future Work
Problem Setting • MANET of m mobile devices – {M 1, M 2, …, Mm} • Local relation Ri on each device Mi – <x, y, p 1, p 2, …, pn> • Skyline issued by a device Morg – <id, posorg, d> • id: network id of query originator Morg • pos: position of Morg • d: distance (from pos) of interest
Technical Challenges • Slow and unreliable wireless channels compared to wired connections – To reduce data transferred between devices • Resource-constrained devices – Storage and processing saving techniques on mobile devices
Outline • Introduction • Problem Definition • Skyline Queries in MANETs • Optimizations on Mobile Devices • Experimental Studies • Conclusion
Straightforward Strategy • Query originator Morg – Executing a local skyline query: SKorg – Sends query to other mobile devices – Merges results when receiving them • A mobile device Mi – Executing a local skyline query too – Sends result SKi back to Morg – Instead of sending whole Ri
Discussion • Final skyline result: SK USKi U| • SK≠USKi, SK • FSK = USKi–SK • FSK contains all those tuples that are not in SK but sent between devices • Identify SKi–SK on device Mi – Inspiration of semi-join
Filtering Strategy • Any tuple tpi in SKi–SK is dominated by some tuple(s) tpj in SK • Where to find such tpjs? – Pick from Morg’s local result – Send <id, posorg, d, tpj> as query – Mi filters out tuples using tpj • Which one to pick? – Dominating region
Dominating Region • The ability of tpj to dominate others – Tuple value <pj 1, pj 2, …, pjn> – Data space boundaries • Volume of dominating region – VDRj=∏k(bk-pjk) • Choose from SKorg tpflt with max VDRj p 2 Max corner of data space b 2 Dominating Region – Indep. distribution p j 2 0 tp j 1 p 1 b 1
Dominating Ability • Two hotel relations – Price range (20. . 200) – Smaller rating means better (1. . 10) Hotel Price Rating Hotel h 11 20 7 h 21 60 3 (200 -60)*(10 -3)=980 h 12 40 5 h 22 80 2 (200 -80)*(10 -2)=960 h 13 80 7 h 23 120 1 (200 -120)*(10 -1)=720 h 14 80 4 h 24 140 2 − h 15 100 7 h 25 100 4 − h 16 100 3 Relation R 1 Price Rating Relation R 2 (Morg) VDR
Estimated Dominating Region • Over-estimation – VDRj=∏k(maxk-pjk) – maxk: pre-specified larger value • Under-estimation – VDRj=∏k(hk-pjk) – hk: local maximum
Dynamic Filtering Tuples • Three hotel relations – M 4 -> M 3 -> M 1 VDR 31=980 VDR 41=960 Hotel Price Rating Hotel h 11 20 7 h 31 60 3 h 41 80 2 h 12 40 5 h 32 80 5 h 42 120 1 h 13 80 7 h 33 120 4 h 43 140 2 h 14 80 4 h 15 100 7 h 16 100 3 Relation R 1 Relation R 3 Price Rating Relation R 4 (Morg)
Query Log Mechanism • To avoid the same query more than once on any device Mi • Add a tag cnt to query issued by Morg – <id, cnt, posorg, d, tpflt> • Mi records/checks/updates <id, cnt> – Processes and forwards only cntlog<cnt • cnt can be a byte to save cost – A device can issue 256 queries – Reset after a period, say one day
Outline • Introduction • Problem Definition • Skyline Queries in MANETs • Optimizations on Mobile Devices • Experimental Studies • Conclusion
Dataset Storage • Goals – Space efficient – Local processing efficient • Operations – Spatial extent check • Distinct coordinates – Attribute value comparison • Floats • Duplicates
Hybrid Storage Model • Spatial coordinates – Real values – MBRi(xmax, ymax, xmin, ymin) • Attribute values – Ascending domains Relation R – IDs x y 1. 331 103. 67 – Sort p 1 i p 1 … pn Sorted domains p 1 pn 0 … 2 v 0 1. 329 103. 59 1 … 0 v 1 1. 412 103. 77 2 … j … … … … vk vj 1. 429 103. 95 k … 3 … v 1
Local Skyline Computing • Sptial check – mindist(posorg, MBRi) > d • Skyline computing – Comparison of IDs instead of true values of float type – p 2 to pn only • Update filtering tuple if necessary – Choose the one with larger VDR value
Assembly on Query Originator • When Morg receives SKi from others – Duplication elimination – False positive removing • A simple nested loop is enough – Comparing coordinates • Identify duplicates – Comparing attribute values • Identify false positive reports from both SKorg and SKi
Outline • Introduction • Problem Definition • Skyline Queries in MANETs • Optimizations on Mobile Devices • Experimental Studies • Conclusion
Experiment Parameters Number of mobile device Cardinality of global reln Cardinality of local reln Local storage model 32, 42, …, 102 100 K, 200 K, …, 1000 K 10 K, 20 K, …, 100 K Flat, Hybrid Number of non-spatial attr Non-spatial attri range Spatial extent of global reln Attribute distribution Query distance of interest 2, 3, 4, 5 [0. 0, 9. 9], [0, 1000] 1000 X 1000 Indep. , Anti-Correl. 100, 250, 500
Studies on Local Optimization • HP i. PAQ h 6365 pocket PC – MS Windows Mobile 2003 – 200 MHz TI OMAP 1510 processor – 64 MB SDRAM (55 MB user accessible) • Super. Waba – Java-based open-source platform for PDA and smartphone applications – www. superwaba. org
Time vs Local Cardinality • Flat Storage vs Hybrid Storage • Anti-Correlated vs Independent • HS incurs less processing cost
Time vs Local Dimensionality • Average of costs on both distributions – Coz they are close to each other • HS still performs better
Performance in Simulation • Simulated MANET – Ji. ST-SWANS • A Jave based MANET simulator • http: //jist. ece. cornell. edu/ – Pentium IV desktop PC • MS Windows XP • 2. 99 GHz CPU • 1 GB memory
Settings • Device setting – Data partitioned and allocated to devices using a grid of m 1/2 by m 1/2 – 1 -5 queries per device • MANET settings – Total simulation time: 2 hours – Speed range: 2 unit/s – 10 unit/s – Holding time: 120 seconds – Wireless routing protocol: AODV
Data Reduction Efficiency • Data Reduction Rate – – SKi’ is the local skyline after filtering • Pre-tests in static setting – Forwarding query out recursively – Findings • No significant difference between exact VDR and estimated VDRs • Dynamic filtering is more powerful
Data Reduction Rate
Response Time - BF • Breadth-First query forwarding – Parallel • Time receiving answers from 80% other devices – Cannot ensure all devices are always reachable and available in MANETs M 2 M 1 M 3 Morg M 5 M 4 Query message Result message
Response Time - DF • Depth-First forwarding – Serialized • Query ends when originator finds all neighbors have processed the query M 2 M 3 M 1 M 4 M 5 Morg Query message Result message
Response Time
Query Message Count • Only mobile device number affects the query message count obviously • Better performance of BF is not free
Outline • Introduction • Problem Definition • Skyline Queries in MANETs • Optimizations on Mobile Devices • Experimental Studies • Conclusion
Conclusion • Problem setting – MANET of lightweight devices – Skyline queries with spatial constraints • Solution highlights – Filtering based distributed query processing strategy to reduce communication cost – Specialized local storage and algorithm to speed up local processing – Experimentally verified performance
Q&A Thanks!
Скачать презентацию Skyline Queries Against Mobile Lightweight Devices in MANETs
01075821d14798bbfbefadb4a086e708.ppt