1. A probabilistic approach to building large scale federated systems Francisco Matias Cuenca-Acuna http://www.panic-lab.rutgers.edu/

2. Federated Computing Rising Internet connectivity is driving a new model of federated computing −Computing systems that span multiple organizations −Sharing of resources, including data and services Federated computing appearing at every level −Social group-based sharing −P2P: Gnutella, KaZaA, DirectConnect −Web-based: Ebay, Google groups, Yahoo groups, DMOZ −Scientific computing −Seti@home −Research grids: The European Data Grid −E-commerce −Federated web ecommerce (Amazon WS) −Distributed supply chains (Travelocity, Sabre)

3. The Challenge Federated computing is a natural model for harnessing inherently distributed resources −Consider data generation and storage −Users produce 740TB of information per year 1 −The European Data Grid has 100’s of nodes hosting PB’s of data Challenge: how do we build systems that are −Inherently decentralized −Widely distributed −Widely heterogeneous −Resilient to uncontrolled node behavior 1 Source http://www.sims.berkeley.edu/research/projects/how-much-info/

4. The PlanetP Project Infrastructure support for federated systems −Communication and distributed state maintenance −Loosely replicated state −Global index over all data stores −Global membership −One to many data propagation −Information sharing −Content addressing & ranking of results −Provide predictable data availability −Deployment, monitoring, and management of federated services −Provide a common runtime environment −Self-managing and self-configuring given quality of service goals

5. Approach & Status PlanetP Principles −Autonomous actions −Loosely synchronized global information −Randomized algorithms PlanetP today … −Multidimensional indexed data store −Accommodates communities of 1000’s nodes −Content ranking comparable to centralized text-based solution −~4% loss of recall and precision when compared to centralized TFxIDF implementation −Practical data availability −Environment modeled after Gnutella, avg availability 24%, can achieve 99.9% data availability with 6x excess storage −Successfully help a replicated service adapt to a volatile environment −Maintains a UDDI service running on Planetlab across 100 nodes

6. The PlanetP Architecture Node X Hoarded Set F1F1 F2F2 FiFi Excess Storage FjFj FkFk Global Data Index MembershipInfo. Gossiping Information Search & Ranking

7. Communication and State Maintenance

8. Nodes push and pull randomly from each others −Unstructured communication  resilient to failures −Predictable convergence time Novel combination of previously known techniques −Rumoring, anti-entropy, and partial anti-entropy −Introduce partial anti-entropy to reduce variance in propagation time for dynamic communities −Batch updates into communication rounds for efficiency −Dynamic slow-down in absence of updates to save bandwidth Epidemic Communication   ___

9. [K 1,..,K n ] Local Objects Bloom filter Local Index Global Directory Gossiping [K 1,..,K n ] Local Objects Bloom filter Local Index Global Directory NicknameStatusIPKeys AliceOnline…[K 1,..,K n ] BobOffline…[K 1,..,K n ] CharlesOnline…[K 1,..,K n ] Globally Indexed Data Store Each node maintains a local index of its shared objects −Objects can be accessed through handles or keys −Summarize the set of keys in its index using a Bloom filter The global index is the set of all summaries −Key-to-peer mappings (but don’t know the exact set of keys) −List of online peers NicknameStatusIPKeys AliceOnline…[K 1,..,K n ] BobOffline…[K 1,..,K n ] CharlesOnline…[K 1,..,K n ]

10. Data Propagation Performance Arrival and departure experiment (LAN)Propagation speed experiment (DSL)

11. Searching & Ranking

12. Content-based Searching Parse & index all documents −Extract keywords/terms to use them as references to the document −Keep a per document term count Advertise terms using the indexed data stored −Effectively build a local inverted index Approximate a global inverted index −Split searching in two phases

13. Content-based Searching Query Diane Global Directory [K 1,..,K n ]Gary [K 1,..,K n ]Fred [K 1,..,K n ]Edward [K 1,..,K n ]Diane [K 1,..,K n ] Keys Charles Bob Alice Nickname Bob Fred Local lookup Fred Bob Diane Rank nodes Diane Contact candidates Fred File 3 File 1 File 2 Rank results STOP

14. Results Ranking The Vector Space model −Documents and queries are represented as k-dimensional vectors −Each dimension represents the weight of a term to a document or query −The angle between a query and a document indicates their similarity Weight assignment (TFxIDF) −Use Term Frequency (TF) to weight terms for documents −Use Inverse Document Frequency (IDF) to weight terms for query −Intuition −TF indicates how relevant a document is to a particular concept −IDF gives more weight to terms that are good discriminators between documents

15. Approximating TFxIDF TFxIDF is not well suited to decentralized environments −Requires term to document mappings −Requires a frequency count for every term in the shared collection Instead, use a two-phase approximation algorithm Replace IDF with IPF (Inverse Peer Frequency) −IPF(t) = f(No. Peers/No. Peers with documents containing term t) −Individuals can compute a consistent global ranking of peers and documents without knowing the global frequency count of terms Rank peers using

16. Pruning Searches Centralized search engines have index for entire collection −Can rank entire set of documents for each query In a distributed search, we do not want to contact peers that have only marginally relevant documents −Stop the search after contacting n peers that did not contribute to current top k ranked documents −n needs to be a function of community size and k

17. Evaluation Answer the following questions −What is the efficacy of our distributed ranking algorithm? −What is the storage cost for the globally replicated index? Evaluation methodology −Use a running prototype to validate and collect micro benchmarks (tested with up to 200 nodes) −Use simulation to predict performance on big communities

18. Ranking Evaluation Evaluated for 5 benchmark document collections AP89 collection from TREC −84678 documents, 129603 words, 97 queries, 266MB −Each collection comes with a set of queries and binary relevance judgments We measure recall (R) and precision (P)

19. AP89 Results Results intersection is 70% at low recall and gets to 100% as recall increases To get 10 documents, PlanetP contacted 20 peers out of 160 candidates

20. Size of Global Index TREC collection (pure text) −Simulate a community of 1000 nodes −Distribute documents uniformly −944,651 documents taking up 3GB −16MB of RAM are needed to store the global index −36MB for 5000 peers −This is 0.5% of the total collection size MP3 collection (audio + tags) −Based on Gnutella measurements −3,000,000 MP3 files taking up 14TB −36MB of RAM are needed to store the global index for 5000 peers −This is 0.0002% of the total collection size

21. Automatic replication for availability

22. Increasing Data Availability GOAL: provide predictable data availability in P2P systems −E.g., for file systems, we want to reason about minimum file availability Wide range of node availability −Node MTTF no longer determined by hardware reliability but by users’ on-line behavior −Fixed number of replicas too wasteful −E.g., small number of replicas on highly available nodes equivalent to many replicas on low available nodes −Gnutella span from 0.1% to 100%, with an average of 24% −Also, we need to recreate replicas as nodes join and leave Long term dynamic membership −In fact, a fixed number doesn’t work at all because availability profile will likely change over time

23. Our Approach  Use replication but −Vary number of replicas based on estimated file availability −Take advantage of nodes going offline as opposed to failing −Loosely monitor availability −Use erasure codes to minimize space requirements and spread file to more nodes

24. Internet The Strategy Advertise: - Availability 20% - Files F1, F2 - Fragments Fi, Fj, Fk Node A Global Data Index Hoarded Set F1F1 F2F2 FiFi Excess Storage FjFj FkFk MembershipInfo. Gossiping Node B Global Data Index Hoarded Set F3F3 F4F4 FxFx Excess Storage FyFy FzFz MembershipInfo. Gossiping

25. Internet Node A Global Data Index Hoarded Set F1F1 F2F2 FiFi Excess Storage FjFj FkFk MembershipInfo. Gossiping The Strategy PlanetP Hoarded Set F3F3 F4F4 FxFx Excess Storage FyFy FzFz MembershipInfo. Gossiping Based on Node’s B view of F3: - Pick a random node - Create a new fragment for F3 - Push it  F3F3 Node B Global Data Index Hoarded Set F3F3 F4F4 FxFx Excess Storage FyFy FzFz MembershipInfo. Gossiping

26. Dealing with Decentralization Nodes replicate and evict autonomously All decisions are probabilistic −Weighted by availability estimates Target nodes control their own storage space −Protects system against greedy and faulty nodes Erasure codes plus −Use a modified version of Reed Solomon −Provide a large fragment space −Don’t re-create lost fragments −Prevents duplicates due to autonomous and misinformed decisions

27. Availability-based Replacement Estimating file availability −Probability of finding an online copy or being able to reconstruct the file from the erasure coded fragments Evict fragments of files with “too much” availability −Note that “too much” is in comparison only to files in local excess storage (don’t have to know about all files in system) Why does it work? −Randomized placement decisions  local sample of file availabilities reflect global distribution −This approximation drives space allocation and allows files with insufficient availability to gain fragments

28. Evaluation Evaluate three significantly different environments The file sharing environment −1000 nodes hosting a total of 25000 files −Node availability avg:24%, min:0.1%, 90th perc:75%, max:100% −Target 99.9% availability −10 minute refresh rate Sources −Saroiu et al. (Gnutella, Napster), DirectConnect at Rutgers OMNI −Centralized knowledge with no limitation on replica placement Base −What happens if you do not have availability estimates?

29. Availability Comparison

33. Effect of Av. Based Replacement

34. Content search & replication? When distributing MED across 100 nodes we find −Avg. nodes per term 16 −Avg. artificial replicas per term 7 (for FS) −Worst case, few nodes & lots of replicas

35. Self-managed Federated Services

36. Adaptive Federated Services GOAL: Reduce administration burden −Operator errors account for 19%-33% of total errors −50% of them are due to configuration problems −Federated environments will amplify this trend GOAL: Improve availability & fault tolerance −Automatic reconfiguration, failure masking & av. estimation Distributed runtime for Web Services −Administrators just dictate the policy −They reason about −capacity −availability −privacy issues −Provide self deployment and monitoring −Wrap service replicas with autonomous agents

37. Example run

38. Conclusions Explored infrastructural support for applications running on federated systems −Membership, content addressing & ranking, service management −Scale well to thousands of peers −Extremely tolerant to unpredictable dynamic peer behaviors Gossiping with partial anti-entropy is reliable −Information always propagate everywhere −Propagation time has small variance Distributed approximation of TFxIDF −Within 11% of centralized implementation −Global index << size of data collection −0.5% for 1000 peers sharing TREC

39. Conclusions Practical data availability −We can achieve 3 in spite of low node availability and decentralized environment −CO: 80% avg. av.  1X −FS: 24% avg. av.  6X −WG: 33% avg. av.  9X −Having some global information is critical −But can do quite well with loosely synchronized data Effective service management & monitoring −No. service replicas adapts to −Environmental changes −Workload changes −Application failures −Monitoring agents can operate autonomously −Probabilistic serialization effectively reduces collisions −They advance toward stable solutions

40. The PlanetP Project http://www.panic-lab.rutgers.edu/ Thank you Questions?