1. DONAR Decentralized Server Selection for Cloud Services B96B02016 生化科技四 張煥基 B97901184 電機三 姜慧如 2011.06.21

2. Introduction The trend toward geographically-diverse server placement will only continue and increasingly include smaller enterprises, with the success of cloud-computing platforms like Amazon AWS. These services all need an effective way to direct clients across the wide area to an appropriate service location (or “replica”).

3. 說文解字： Replica Selection Geo-replicated services need an effective way to direct client requests to a particular location, based on performance, load, and cost.

4. Replication Selection 兩大主 流 central coordination distributed heuristics(DONAR) 優點 (1) reliability (2) security 同時解決 (1) client performance (2) server load 缺點 (1) single point of failure (2) attractive target for attackers (3) overhead (4) less responsive to sudden changes (5) scalability limitations (1) nodes over-react based on their own local information (2) the system does not balance replica load effectively

5. Replica-Selection System 必備特質 [1] Expressive Customers should have a sufficiently expressive interface to specify policies based on (some combination of) (1) performance, (2)replica load, and (3) server and bandwidth costs. [2] Reliable The system should offer reliable service to clients, as well as stable storage of customer policy and replica configuration data. [3] Accurate Client requests should be directed to the service replicas as accurately as possible, based on the customer’s replica-selection policy. [4] Responsive The replica-selection system should respond quickly to changing client demands and customer policies without introducing instability. [5] Flexible The nodes should support a variety of replica- selection mechanisms [6] Secure Only the customer, or another authorized party,should be able to create or change its selection policies.

6. 本篇主角： DONAR This paper presents DONAR, a distributed system that can offload the burden of replica selection, while providing these services with a sufficiently expressive interface for specifying mapping policies.

7. 1.2 Decentralized Replica- Selection System mapping node 的任務 (1) direct its clients (2)adapt to changing conditions

8. Roadmap section 2 Simple and expressive interface for customer policies section 3 Stable, efficient, and accurate distributed replica-selection algorithm section 4 Scalable, secure, reliable, and flexible prototype system section 5 Experiments in Section 5 evaluate both our distributed algorithm operating at scale and a small-scale deployment of our prototype system section 6 compares DONAR to related work section 7 concludes

9. 2.1 Customer Goals Customers use DONAR to optimally pair clients with service replicas minimize the network latency balance load across all replicas billing costs

10. 2.2 Application Programming Interface create a DONAR service s = create () add a replica instance i = add (s, repl, ttl) time-to-live period (ttl) set split weightset (s, i, w i, ε i ) set bandwidth capset (s, i, B i ) match a client-replica pair match (s, clnt, i) prefer a particular replicapreference (s, clnt, i) remove a replica instance remove (s, i)

11. Roadmap section 2 Simple and expressive interface for customer policies section 3 Stable, efficient, and accurate distributed replica-selection algorithm section 4 Scalable, secure, reliable, and flexible prototype system section 5 Experiments in Section 5 evaluate both our distributed algorithm operating at scale and a small-scale deployment of our prototype system section 6 compares DONAR to related work section 7 concludes

12. 3.1 Global Replica-Selection Problem 若想提高網路效能，就得以 accurate load distribution 為代價。 Our goal is to minimize this performance penalty

13. 3.2 Distributed Mapping Service 每個 mapping node 各有其負責的 clients The node maps the client to a replica, and returns the result to that client.

14. 3.2 Distributed Mapping Service 所有 clients 的 traffic, node n 所 佔的比例 mapping node n 所有的 traffic 中，從 client c 而來的 比例 所有從 client c 而來 ，經過 mapping node n 的 traffic ， 流入 replica i 的比例, i.e., ∑ i R nci = 1

15. 3.3 Decentralized Selection Algorithm optimization decomposition : 藉由 algorithmic iterations, 讓 local decisions converge to the global optimum. global performance local client performance

16. The optimization of local performance. 每一個 mapping node 以 client population ＆ replica 上的 load term 來 optimizes local performance the unit price of violating the constraint. True proportion of requests directed to replica i

17. local replica selection 某特定 mapping node n, 將其所負 責之所有 clients 的 traffic 引到 replicas 所需的 performance penalty. load n = load, ∀ n ；超出預期流量的罰款。 ( 以一個 mapping node 的視野看世界 )

18. The core components of the algorithm are the local updates by each mapping node, and the periodic updates of replica prices.

19. overhead centralized solution distributed solution Each node needs to share its mapping decisions of size and each replica’s price λ i needs to be known by each node. This implies messages, each of size computational complexity is of size 每個 mapping node 會有 |N-1| 個 mapping node 的鄰居。他們會告訴 此 mapping node 個與 replica mapping 有關的消息。 個 replicas 會告訴每一個 mapping node 他們的 price

20. DONAR’s system design

22. Distribution optimization---tracking requests geographically A group of similarly located end-hosts.

23. Distributed Optimization-- Tracking Requests Geographically

24. Distributed Optimization-- Exponentially weighted moving average

25. Distributed Optimization-- Known cost assumption

26. DONAR’s System Design

27. Decomposed Local Problem For Some Node (n*)

28. DONAR Algorithm

33. Better!

34. DONAR’s System Design

35. Protocol-level mechanisms for Wide-area replica selection

36. Data Retrieving Steps

37. Data Retrieving Steps(cont.)

38. DONAR’s System Design

39. Secure Registration and Dynamic Updates

40. DONAR’s System Design

41. Distributed Data Storage

42. CRAQ (Chain Replication with Apportioned Queries) “while maintaining the strong consistency properties of chain replication, provides lower latency and higher throughput for read operations by supporting apportioned ( 分攤 ) queries: that is, dividing read operations over all nodes in a chain, as opposed to requiring that they all be handled by a single primary node.”

43. DONAR’s System Design

44. IP Anycast

45. Software Architecture

46. Results: DONAR Curbs Volatility

47. Results: DONAR Minimizes Distance

48. Conclusions

49. Thank You