1. Querying the Internet with PIER (PIER = Peer-to-peer Information Exchange and Retrieval) Ryan Huebsch † Joe Hellerstein †, Nick Lanham †, Boon Thau Loo †, Scott Shenker ‡, Ion Stoica † p2p@db.cs.berkeley.edu † UC Berkeley, CS Division ‡ International Computer Science Institute, Berkeley CA VLDB 9/12/03

2. 2 What is Very Large? Depends on Who You Are Single Site Clusters Internet Scale 1000’s – Millions Distributed 10’s – 100’s Challenge: How to run DB style queries at Internet Scale! Database CommunityNetwork Community

3. 3 What are the Key Properties? Lots of data that is: 1.Naturally distributed (where it’s generated) 2.Centralized collection undesirable 3.Homogeneous in schema 4.Data is more useful when viewed as a whole This is the design space we have chosen to investigate.

4. 4 Who Needs Internet Scale? Example 1: Filenames Simple ubiquitous schemas: Filenames, Sizes, ID3 tags Born from early P2P systems such as Napster, Gnutella, AudioGalaxy, etc. Content is shared by “normal” non-expert users… home users Systems were built by a few individuals ‘in their garages’  Low barrier to entry

5. 5 Example 2: Network Traces Schemas are mostly standardized: IP, SMTP, HTTP, SNMP log formats Network administrators are looking for patterns within their site AND with other sites: DoS attacks cross administrative boundaries Tracking virus/worm infections Timeliness is very helpful Might surprise you how useful it is: Network bandwidth on PlanetLab (world-wide distributed research test bed) is mostly filled with people monitoring the network status

6. 6 Our Challenge Our focus is on the challenge of scale: Applications are homogeneous and distributed Already have significant interest Provide a flexible framework for a wide variety of applications Other interesting issues we will not discuss today: Heterogeneous schemas Crawling/Warehousing/etc These issues are complementary to the issues we will cover in this talk

7. 7 Four Design Principles (I) Relaxed Consistency ACID transactions severely limits the scalability and availability of distributed databases We provide best-effort results Organic Scaling Applications may start small, without a priori knowledge of size

8. 8 Four Design Principles (II) Natural habitat No CREATE TABLE/INSERT No “publish to web server” Wrappers or gateways allow the information to be accessed where it is created Standard Schemas via Grassroots software Data is produced by widespread software providing a de-facto schema to utilize

9. Physical NetworkOverlay Network Query Plan Declarative Queries

10. 10 Recent Overlay Networks: Distributed Hash Tables (DHTs) What is a DHT? Take an abstract ID space, and partition among a changing set of computers (nodes) Given a message with an ID, route the message to the computer currently responsible for that ID Can store messages at the nodes This is like a “distributed hash table” Provides a put()/get() API Cheap maintenance when nodes come and go

11. 11 Recent Overlay Networks: Distributed Hash Tables (DHTs) (16,16) (16,0) (0,16)(0,0) Data Key = (15,14)

12. 12 Recent Overlay Networks: Distributed Hash Tables (DHTs) Lots of effort is put into making DHTs better: Scalable (thousands  millions of nodes) Resilient to failure Secure (anonymity, encryption, etc.) Efficient (fast access with minimal state) Load balanced etc.

13. 13 PIER’s Three Uses for DHTs Single elegant mechanism with many uses: Search: Index Like a hash index Partitioning: Value (key)-based routing Like Gamma/Volcano Routing: Network routing for QP messages Query dissemination Bloom filters Hierarchical QP operators (aggregation, join, etc) Not clear there’s another substrate that supports all these uses

14. 14 Metrics We are primarily interested in 3 metrics: Answer quality (recall and precision) Bandwidth utilization Latency Different DHTs provide different properties: Resilience to failures (recovery time)  answer quality Path length  bandwidth & latency Path convergence  bandwidth & latency Different QP Join Strategies: Symmetric Hash Join, Fetch Matches, Symmetric Semi-Join, Bloom Filters, etc. Big Picture: Tradeoff bandwidth (extra rehashing) and latency

15. 15 Experimental Setup A simple join query SELECTR.pkey, S.pkey, R.pad FROMR, S WHERER.num1 = S.pkey AND R.num2 > constant1 AND S.num2 > constant2 AND f(R.num3, S.num3) > constant3 Table R has 10 times more tuples than S num1, num2, num3 are uniformly distributed R.num1 is chosen such that 90% of R tuples have one matching tuple in S, while 10% have no matching tuples. R.pad is used to insure all result tuples are 1KB Symmetric Hash Join

16. 16 Does This Work for Real? Scale-up Performance (1MB source data/node) Simulation Real Network

17. 17 Current Status of PIER PIER currently runs box-and-arrow dataflow graphs We are working on query optimization GUI (called Lighthouse) for query entry and result display Designed primarily for application developers User inputs an optimized physical query plan

18. 18 Future Research Routing, Storage and Layering Catalogs and Query Optimization Hierarchical Aggregations Range Predicates Continuous Queries over Streams Sharing between Queries Semi-structured Data

20. Bonus Material

21. 21 Symmetric Hash Join (SHJ)

22. 22 Fetch Matches (FM)

23. 23 Symmetric Semi Join (SSJ) Both R and S are projected to save bandwidth The complete R and S tuples are fetched in parallel to improve latency