1. Ad-hoc Distributed Spatial Joins on Mobile Devices Panos Kalnis, Xiaochen Li National University of Singapore Nikos Mamoulis The University of Hong Kong Spiridon Bakiras Hong Kong University of Science and Technology

2. Motivation  Users are equipped with a mobile device (eg. PDA)  Ad-hoc spatial queries  Combine data from remote servers Hotels Restaurants “Find hotels which are within 500m of a seafood restaurant”  Servers do not collaborate with each other  The query is executed on the mobile device

3. Mediators?  Services may only allow end-user connections (eg., subscribers only)  Access through mediators may be more expensive  Requests are ad-hoc; existing mediators may not support them Hotels Restaurants Mediator

4. Cost  Telecommunication companies typically charge by the bulk of transferred data (eg. GPRS), instead of connection time.  Goal: Minimize the amount of transferred data.

5. Solution  Ask aggregate queries to estimate the data distribution (i.e., statistics)  Partition the space recursively to achieve sub-linear transfer cost  Choose the physical operator indepen- dently for each partition

6. Related Work  Hash-based methods (eg. PBSM): require all data to be transferred  R-tree based methods (eg., [Tan et.al, TKDE, 2000]): require access to internal index  Mediators : HERMES : Statistics from previous queries DISCO, Garlic : Statistics during initialization Tuckila : Optimize parts of the execution tree

7. Operators  WINDOW query: return all objects intersecting a window w  COUNT query: return the number of objects intersecting w  ε-RANGE query: return all objects within range ε from a point p NO access to the internal indices! ε w p

8. Query Types  Intersection Join Find hotels which are inside parks  E-range Join Find restaurants which are within 500m of a hotel  Iceberg Semi-join Find hotels which are close to at least 3 restaurants ε

9. H ash B ased S patial J oin Each partition must fit in memory

10. Recursive evaluation Retrieve statistics for each subpart

11. Inefficient HBSJ

12. N ested L oop S patial J oin Recursive HBSJ : 4 QRY + 2 RCV + 5 RCV NLSJ : 2 RCV + 2 SND + 2 RES

13. Inefficient NLSJ

14. Cost Model  TCP/IP: MTU = MSS + B H  c1: download |R W | objects from R and |S w | objects from S and join them on the PDA  C2,3: download |R W | objects from R, send them as window queries to S and retrieve the results  c4: repartition w, retrieve detailed statistics and apply the algorithm recursively

15. UpJoin (Uniform Partition Join) Decide if datasets are uniform If HBSJ is cheaper and both datasets are uniform then perform HBSJ If NLSJ is cheaper and the largest dataset is uniform then perform NLSJ Else repartition

16. Uniformity check Dw Dw’0Dw’1 Dw’3Dw’2 % variation from uniform distribution  Note: UpJoin will not repartition if the cost for retrieving statistics is larger than the cost of joining

17. Inefficient UpJoin

18. SR-Join (Similarity Related Join) Area % variation of density  Identify dense and sparse quadrants  If the distribution is similar then apply HBSJ or NLSJ  Else repartition X X X X

19. Experimental setup  Implementation Server: Unix Client: HP-Ipaq PDA (WiFi network, 400MHz RISC CPU, 64MB RAM, Windows Pocket PC)  Datasets: Synthetic: 1K – 10K points, varying skew Real: Roads and railways of Germany

20. Setting the parameters α (for UpJoin)ρ (for SR-Join) Uniform

21. Real Dataset Uniform

22. Comparison with SemiJoin SemiJoin: Use intermediate levels of R-Tree index We cannot use it in practice, because we cannot access the index Uniform

23. Conclusions  Distributed spatial joins on mobile devices  No mediator – non collaborative servers – limited set of supported operators  Two algorithms UpJoin SRJoin Both estimate the datasets’ distribution  Future work Support multi-way spatial joins Improve the accuracy of the cost model

24. Questions?