1. Epidemic Techniques Chiu Wah So (Kelvin)

2. Database Replication Why do we replicate database? – Low latency – High availability To achieve strong (sequential) consistency on replicated database. Not scalable. – One primary database – Quorum system (contact over half of replicas)

3. Database Replication To scale and have high availability, we need a weaker consistency model – Eventual consistency: If all updating stops then eventually all replicas will converge to the identical values. The two papers talk about how to use epidemic techniques to achieve eventual consistency to scale.

4. Epidemic Techniques for replicated database Epidemic Algorithms for Replicated Database Maintenance – Look at different epidemic algorithms to reduce bandwidth consumption to maintain replicated database Astrolabe – Scalable and Robust information management system.

5. Motivation on the first paper Clearinghouse service maintains translations from names to machine addresses. (like DNS) Problem: Using direct mail and anti-entropy, too much traffic to maintain consistency between highly replicated servers. Some key links are overloaded. Look at techniques to reduce bandwidth: rumor spreading and spatial distributions.

6. Direct mail Direct mail: each server sends update to all other servers. Advantage – Easy to implement – Good enough for small and static servers Disadvantage: – Not scale (O(n) message for each update) – Updates may get lost.

7. Anti-entropy Servers pick random server and resolve differences. 3 ways to resolve differences: push, pull, and push- pull.

8. Anti-entropy Example

9. Push

10. Pull

11. Push-Pull

12. Anti-entropy Average time Average converge in O(log(n)) steps Pull, push-pull Push

13. Anti-entropy (2) Very expensive to send the whole database across network to compare Some techniques for optimizing comparing bandwidth – Compute Checksum – Exchange list of of recent updates. Then apply the update and compute checksum – Exchange updates in reverse chronological order until checksums agree Still too much bandwidth…..

14. Rumor spreading Main idea: send out updates randomly. Instead of comparing whole database. Three states: susceptible, infective, and removed. Initially all servers are susceptible Once server has a rumor (infective), and then pick a random server to send the rumor. With probability 1/k, the server loses interest (removed) to spread rumor

15. Rumor spreading (2) But maybe not every server got the rumor. – With probability of remaining susceptible after the epidemic finishes: Run anti-entropy infrequently to make sure every server gets the update.

16. Three goals in rumor spreading Low Residue: the probability of remaining susceptible when the epidemic finishes Low Traffic: total traffic sent per site Low Delay: Average time and the last time between the injection of an update and the arrival of update.

17. Variations in rumor spreading Many variations in rumor spreading. – Blind with Coin vs Feedback with Counter – Push vs Pull – Increase the smaller counter of the two – Connection limit – Hunting

18. Feedback Counter vs Blind Prob. Feedback and Counter Blind and Probability

19. Deletion and Death Certificates Simple solution: death certificates and store for a fixed threshold of time 2 nd solution: dormant death certificates. Use two threshold time, and some servers keep it longer. 2 different timestamp: original timestamp and reactivation timestamp.

20. Motivation of Spatial Distributions Network is not uniform. Certain key links in the network are overload. – Transatlantic links about 80 conversations, but on average conversations per link is 6. Therefore, we should favor nearby neighbors.

21. Spatial Distributions Each servers, sort the list of sites by distance from s. Select anti-entropy exchange partners from the sorted list according to a function f(i), i = index on the sorted list. We can use f(i) = i^(-a), where a is the parameter for tuning spatial distribution.

22. Spatial Distribution

23. Next Paper: Astrolabe The first paper talks about how to use rumor spreading and spatial distribution to reduce bandwidth. But the storage grows O(n) and total bandwidth taken up by gossip grows O(n^2) We need a more scalable solution.

24. Astrolabe Scalable and Robust information management system. Monitors the dynamically changing state of a collection of distributed resources. Reports summaries of this information to users.

25. Four design goals Scalability: scale through its zone hierarchy. Information is summarized before exchanges. Flexibility: easy to install new aggregated function in a form of SQL aggregation query Robustness: randomized peer-to-peer approach to exchange information. Security: use signed certificates.

26. Examples: uses of astrolabes Peer-to-Peer Caching of Large Objects Peer-to-Peer Data Diffusion Publish-Subscribe Synchronization

27. Structure of Astrolabe Structure of Astrolabe’s zones can be viewed as a trees. Leaves of this tree are hosts. Each hosts run an astrolabe agent.

28. Astrolabe Detail Each agent is a virtual database. Each agent has a path name. (For example: /USA/Cornell/pc3) Each agent contains information, called MIB, for all the ancestor zone (For example, it contains /, /USA, /USA/Cornell) Each ancestor MIB is generated using aggregation for scalability, instead of having O(n) entries.

29. Astrolabe Detail (2) Each zone can be viewed as relational table of the attributes of its child zone. How do we gather or generate the information in the zone relational table? Two ways: If the agent is in the zone, use aggregation to construct the MIB for the zone. Otherwise, gossip for the information. Therefore, MIB for internal zones has to be small in order to scale.

30. Aggregation Aggregation Function Certificates contain information on how to collect and aggregate attributes of child zone MIBs into entries for internal zone MIBs. – Programmed in SQL-like language – Propagates by two ways: copying to parent (propagates like other normal attributes), and look for new AFC from its ancestor zone

31. Aggregation (2) Here are the SQL aggregation functions that are provided by Astrolabe.

32. Gossip Each zone has a small set of addresses for representative agents. Representative agents are computed using an aggregation function, such as using load and longevity. An agent gossips on behalf of those zones for which it is a representative.

33. Gossip (2) Periodically, the agent picks one of the child zones, and talks to one of the contact agents. (anti-entropy) Then, it sends all the child zones at that level, and does the same thing for the higher levels in the tree up until the root level. Then the two agents can compare which entries are newer and keep them.

34. Example of gossip (taken from ken slides) NameTimeLoadWeblogic?SMTP?Word Version swift2003.67016.2 falcon19762.7104.1 cardinal22013.5116.0 NameTimeLoadWeblogic?SMTP?Word Versio n swift20112.0016.2 falcon19711.5104.1 cardinal20044.5106.0 swift.cs.cornell.edu cardinal.cs.cornell.edu

35. Example of gossip (2) NameTimeLoadWeblogic?SMTP?Word Version swift2003.67016.2 falcon19762.7104.1 cardinal22013.5116.0 NameTimeLoadWeblogic?SMTP?Word Versio n swift20112.0016.2 falcon19711.5104.1 cardinal20044.5106.0 swift.cs.cornell.edu cardinal.cs.cornell.edu swift20112.0 cardinal22013.5

36. Example of gossip (3) NameTimeLoadWeblogic?SMTP?Word Version swift20112.0016.2 falcon19762.7104.1 cardinal22013.5116.0 NameTimeLoadWeblogic?SMTP?Word Versio n swift20112.0016.2 falcon19711.5104.1 cardinal22013.5106.0 swift.cs.cornell.edu cardinal.cs.cornell.edu

37. NameLoadWeblogic?SMTP?Word Version … swift2.0016.2 falcon1.5104.1 cardinal4.5106.0 NameLoadWeblogic?SMTP?Word Version … gazelle1.7004.5 zebra3.2016.2 gnu.5106.2 NameAvg Load WL contactSMTP contact SF2.6123.45.61.3123.45.61.17 NJ1.8127.16.77.6127.16.77.11 Paris3.114.66.71.814.66.71.12 San Francisco New Jersey SQL query “summarizes” data Dynamically changing query output is visible system-wide

38. Membership When an agent has not seen an update for a zone from a particular representative for some time Tfail. Remove its MIB. Connect different pieces of the trees and add in new machines – IP multicast – Broadcast – Relatives Administrators responsible for configuring the system by assigning zone names.

39. Communication Through Http and UDP(need to fragment the messages into more than one UDP packets) If there is firewall, – Use ALG in core internet or an astrolabe agent in core internet.

40. Security Each zone is a management unit. Children have a way to override policy enforced by parents. Each zone: 2 pairs of key, CA and zone keys – Zone certificate – MIB certificate – Aggregation function certificate – Client certificate

41. Related work Directory Services (Clearinghouse, Bayou, Globe) Network Monitoring Event Notification Sensor Networks Peer-to-peer routing

42. Measurement on expected # rounds

43. Measurement on expected # rounds (2)

44. Measurement on Latency RealSimulation

45. Conclusion ??