1. Hybrid Keyword Search across Peer-to-Peer Federated Data PhD Dissertation Defense Florida State University Jungkee (Jake) Kim

2. Motivation Internet Where is the Information?

3. Outline Two Typical Search Paradigms Problem Statements of Current Approaches Hybrid Keyword Search Hybrid Search on Distributed Databases Hybrid Search across Peer-to-Peer Federated Databases

4. Two Typical Search Paradigms Searching over structured data Relational Databases Searching over unstructured data Information Retrieval Internet Environment Semistructured Data – XML Keyword Search in DB Web Search Engines – Technologies from Information Retrieval Hybrid Keyword Search ?

5. Current Approaches – Keyword-only Search Web Search Engines Web crawlers visit Web pages and collect the keyword based text indexes. Web crawlers visit Web pages and collect the keyword based text indexes. Fast information retrieval Fast information retrieval Keyword Search in databases Web integration on legacy DBMS Web integration on legacy DBMS Dynamic Web publication through embedded DB Dynamic Web publication through embedded DB Easy to use without knowledge of DB schema Easy to use without knowledge of DB schema

6. Problems of Current Approaches – Keyword-based Web Search Engines Can not collect every connected resource Can not collect every connected resource Query results are often unrelated Query results are often unrelated Keyword Search in Databases Losing the inherent meaning of the schema Losing the inherent meaning of the schema Query results are not based on semantic schema Query results are not based on semantic schema

7. Current Approaches – Semantic Semantic Web Multiple relation links with directed labeled graphs and machines can understand the relationship between different resources Multiple relation links with directed labeled graphs and machines can understand the relationship between different resources Describes metadata about resources Describes metadata about resources To represent the relations of the objects on the Web; the object terms defined under a specific description – an Ontology To represent the relations of the objects on the Web; the object terms defined under a specific description – an Ontology

8. Problems of Current Approaches – Semantic Web Ontology design is sophisticated Lack of unified definition * * Limited adoption

9. Our Approach Hybrid search mechanisms – Semantic metadata + Keyword search Semantic Solution Semantic Web might be better than Hybrid search Hybrid search must be better than Web search engines Simplicity Hybrid search is simpler than Semantic Web

10. Hybrid Keyword Search Service A search service fetches target information data against a search query. Unstructured data A file containing data – MS Word, PDF, PS documents Metadata: Structured or semistructured data – XML We utilized an XML-enabled relational DBMS and a native XML DB along with a text search library (Apache Xindice + Jakarta Lucene) to address the search against metadata and text.

11. How to Combine? (1) Two entities and a relationship in relational DBMS We can obtain the hybrid search result using a nested subquery

12. How to Combine? (2) A hash table is used for joining search results in non- DBMS based system (Apache Xindice + Lucene)

13. Local Query Processing – XML (1) XML-enabled RDB DBLP XML record DBLP XML record (1,000 – 10,000) (1,000 – 10,000) Non indexed matches except year match bound by the number of matches. Non indexed matches except year match bound by the number of matches. Combined query time depends on # of year query results Combined query time depends on # of year query results Average XML Query Time

14. Local Query Processing – XML (2) Apache Xindice DBLP XML record (1,000 – 10,000) Indexed approximate matches for text elements in XML instances as bad as non- indexed queries Exact matches bound by the number of matches. Average XML Query Time

15. Local Query Processing – Hybrid (1) Hybrid search query performance measurement XML-enabled RDB XML-enabled RDB For 100,000 XML instances and 100,000 text documents For 100,000 XML instances and 100,000 text documents Small result set: 4 XML and a keyword matches Small result set: 4 XML and a keyword matches Large result set: 7,752 XML and 41,889 documents Large result set: 7,752 XML and 41,889 documents MetadataAuthorYear (Nested subquery) Year (Hash table) FewKeywords0.04Sec. 82.9 Sec. 5.70 Sec. ManyKeywords0.48Sec. Half hour 6.96 Sec.

16. Local Query Processing – Hybrid (2) Hybrid search query performance measurement Apache Xindice + Jakarta Lucene For 10,000 XML instances and 10,000 text documents Small result set: 2 XML and a keyword matches Large result set: 192 XML and 4,562 documents

17. Discussion – Local Hybrid Search XML-enabled RDB provides proper response except some extreme query loads. A native XML DB (Apache Xindice) had very limited scalability. (No accurate query result over 16,000 XML instances) We will generalize hybrid search to a distributed environment.

18. Hybrid Search on Distributed Databases Data Independence: logically and physically independent; the same schema – no change, data encapsulation in each machine Network Transparency: depends on MOM or P2P framework No replication – restricted to a computer cluster Fragment: full partition; horizontal fragmentation The query result for the distributed databases is the collection of query results from individual database queries.

19. Scalable Hybrid Search Architecture on DDBS Search Service Message Broker Client Search Service Search Service Subscriber for a query topic Publisher for a temporary topic Publisher for a query topic Subscriber for a temporary topic Query Message Query Message Result Message Result Message Client

20. Cooperating Broker Network Distributed Databases based on NaradaBrokering Network

21. Query Processing – DDBS (1) 100,000 XML and 100,000 Documents in 8 machines – 12,500 each Few keyword match (1-3) on 1 machine only RDB – 0.04 Sec. for few keyword match Avg. response time for an author exact match query over 8 search services

22. Query Processing – DDBS (2) 100,000 XML and 100,000 Documents in 8 machines – 12,500 each RDB – half hour or 6.96 Sec. (Hash table) Avg. response time for a year match query over 8 search services

23. Data Integration Hub Partial integration – possible method to increase the data portion queried c.f. Supernode in P2P We designed a partial integration architecture through a message-oriented middleware – the NaradaBrokering system NaradaBrokering system JMS compliant topic-based communication JMS compliant topic-based communication Scalability by brokers hierarchical connection Scalability by brokers hierarchical connection Passive queries / Static binding Passive queries / Static binding We attached a RDBMS to store the metadata and index the contents of the data

24. Architecture of Data Integration Hub

25. Coupling vs. Scalability From ICDE 2002 Tutorial

26. Query Propagate and Results back on a P2P Network

27. Peer group architecture of the P2P Search

28. Performance Test for Peer Group Communication (JXTA) ….. Subnet ASubnet BSubnet C Client PeerRendezvous Peer Search Service Peers GroupPropagation GroupPropagation Point-to-point Pipe Connection

29. Performance for Group Peer Communication – 1 Peer per Node Average Response Time for a Query

30. Performance for Group Peer Communication – Multiple Peers per Node Allowed (1) Average Response Time for a Query with Multiple Peers per Node Allowed

31. Performance for Group Peer Communication – Multiple Peers per Node Allowed (2) Message Response Time for 32 Group Peers

32. Related Works (1) Distributed lookup in routing to reduce the unnecessary communications Distributed Hash Table (DHT) – Chord, CAN, Pastry, and Tapestry Distributed Hash Table (DHT) – Chord, CAN, Pastry, and Tapestry JXTA: DHT + multiple random walks JXTA: DHT + multiple random walks Look up peers based on reputation Hristidis et. al. – Exploiting a context on existing RDBMS with reducing the schema loss of Keyword Search in DB

33. Related Works (2) MethodMetadata(XML)ContentsNote PlanetPNoYesGossiping Thousands peers ODDISEANoYes Dist. Global index Pastry Galanis and et al. YesNo Dist. Directories Chord, Thousands XRANKYesYes (in XML) No P2P

34. Conclusion We addressed the semantic loss of keyword-only search while remaining a simpler solution than the Semantic Web Low cost scalability over heterogeneous resource through customized overlay networks A practical bridging role on the road towards the ideal of information represented by Semantic Web?

35. Contributions Demonstration of a hybrid search – combining metadata search with a keyword search over unstructured context data A way to increase locality and integrate several dispersed resources through a data integration hub Extension of the scalability of a native XML database and performance improvement for some queries compared to those on a single machine Generalization of our hybrid search architecture on potentially more scalable P2P overlay network

36. Publications J. Kim and G. Fox. Scalable Hybrid Search on Distributed Databases. Accepted for presentation in 3rd International Workshop on Autonomic Distributed Data and Storage Systems Management (ADSM) in conjunction with ICCS, To appear in Lecture Notes in Computer Science. May, 2005. J. Kim and G. Fox. A Hybrid Keyword Search across Peer-to-Peer Federated Databases. In Proceedings of 8th East-European Conference on Advances in Databases and Information Systems (ADBIS), September, 2004. J. Kim, O. Balsoy, M. Pierce, and G. Fox. Design of a Hybrid Search in the Online Knowledge Center. In Proceedings of IASTED International Conference on Information and Knowledge Sharing, November, 2002. G. Aydin, H. Altay, M. S. Aktas, M. N. Aysan, G. Fox, C. Ikibas, J. Kim, A. Kaplan, A. E. Topcu, M. Pierce, B. Yildiz, and O. Balsoy. Online Knowledge Center Tools for Metadata Management. Technical report, DoD HPCMP Users Group Meeting, June, 2003. O. Balsoy, M. S. Aktas, G. Aydin, M. N. Aysan, C. Ikibas, A. Kaplan, J. Kim, M. Pierce, A. Topcu, B. Yildiz, and G. Fox. The Online Knowledge Center: Building a Component Based Portal. In Proceedings of the International Conference on Information and Knowledge Engineering, June, 2002. G. Fox, S. Ko, M. Pierce, O. Balsoy, J. Kim, S. Lee, K. Kim, S. Oh, X. Rao, M. Varank, H. Bulut, G. Gunduz, X. Qiu, S. Pallickara, A. Uyar, and C. Youn. Grid services for earthquake science. Concurrency and Computation: Practice and Experience, 14:371---393, May---June 2002.