1. 1 Content Distribution Networks

2. 2 Replication Issues Request distribution: how to transparently distribute requests for content among replication servers. Server selection: how to select a server replica for a given request. Content placement: how to decide how many replicas of given content to have and on which servers to place these replicas.

3. 3 Request distribution Transparent replication: techniques that require no user involvement or even awareness of the underlying replication scheme. Clients use a single logical name when requesting content. Basic issue: redirection of logically identical requests to distinct servers.

4. 4 Request Redirection At the client At the intermediate proxies At the DNS system At the primary origin server Somewhere in the network

5. 5 Content-blind request distribution with Full Replication Client redirection Redirection by a balancing switch Redirection through DNS servers anycast.

6. 6 Client redirection (1/2) Use the Web client’s DNS.

7. 7 Client redirection (2/2) Client redirection at the client. Use Web site DNS to return a list of server replicas. The entire list is passed to the client by the client DNS. Client selects server replica to contact.

8. 8 Advantages of client schemes Client can make the server selection. Client can compare routing paths, measure response times, packet loss rates or other metrics. Does not interfere with the DNS caching. Problem: server selection in performed by the client (sub-optimal selection of servers).

9. 9 Balancing switch redirection Hardware-based solution to redirect requests. “Balancer” is placed in front of a server farm. Balancer modifies addresses.

10. 10 Switch-based redirection L3, L4 switch: IP balancing Uses standard clients, DNS and Web servers. Low geographical scalability. IBM Network dispatcher.

11. 11 Web Site DNS redirection Many DNS implementations allow the Web site DNS to map a host domain name to a set of IP addresses and chose one of them for every query.

12. 12 Web site DNS redirection: pros and cons Scales well geographically. No modifications of the client or DNS are needed. DNS response caching complicates the management of caching replicas and request distribution.

13. 13 Anycast Multiple physical servers use a single IP addr called anycast addr. Each server advertises both the anycast addr and its regular addr. Routers build paths that lead to the nearest anycast member-server.

14. 14 Anycast

15. 15 Content-blind request distribution with partial replication Servers may have the entire replica of the web site. Full replication is often used but imposes considerable overhead. Typically only a small fraction of content is responsible for most of the requests. Partial replication.

16. 16 Surrogates as server replicas Content requests are distributed among surrogates that are distinct from the origin servers. A surrogate can fulfill the posed request. Otherwise the object is retrieved from the origin server. DNS redirection is often used for distributing requests among the surrogates.

17. 17 Partial replication through surrogates

18. 18 Back-end distributed file systems Use a distributed file system to allow each replication server access the same, shared file set.

19. 19 Content Delivery Networks CDN: agents of content providers A CDN signs up individual content providers for scalable content delivery and delivers their content to any client that accesses the respective Web sites. CDN customers, CDN clients. The latter download content from the CDN that the former provide to the CDN.

20. 20 Benefits to content providers Global reach: A CDN serves content from multiple CDN servers deployed around the globe. By signing up with a CDN, a content provider gains instant presence around the globe at virtually no upfront cost. Flash event protection: Sometimes a Web site experiences a sudden increase in demand called a flash event. A CDN offers an easy way to prepare for a predictable flash event or to protect from an unforeseen one.

21. 21 CDN benefits A CDN is an intermediate layer of infrastructure between origin servers and clients (middleware). A CDN can achieve scalable content delivery by distributing load among its servers, by serving client requests from servers that are close to requesters, and by bypassing congested network paths. CDN infrastructure is shared among multiple content provider sites. CDN has close relationship with the underlying networks. A technical consequence of shared infrastructure is that CDNs must implement a mechanism for finding the origin server for a given piece of content. close relationship with underlying networks: CDNs place their servers within PoPs or backbone nodes of ISPs.

22. 22 Types of CDNs Classification: relationship to ISPs, relationship to customers, mechanisms for delivering requests to CDN servers. Multi ISP, single ISP. CDN with hosting service (CDN servers for relaying content from origin servers and the origin servers themselves, customer maintains a staging server). Relaying CDNs: origin servers remain external to the CDN.

23. 23 CDN types

24. 24 Relaying CDN: 1 st hit @ origin

25. 25 Relaying CDN: 1 st hit @ CDN

26. 26 Request delivery mechanisms DNS outsourcing. Customer delegates the DNS service for its domain to the CDN. Customer’s DNS replies (when queried) with a point to the CDN DNS. CDN DNS resolves query to one of the CDN servers.

27. 27 Relaying CDNs with 1 st hit @ origin Embedded urls that the CDN is supposed to deliver use host names that are controlled by the CDN DNS. Alternatively, the embedded content uses domain names that belong to the CDN. Method adopted by Akamai.

28. 28 Comparing CDN types Origin-first CDNs require that embedded urls use distinct domains from container pages. Relative urls are avoided. CDN-first CDNs allow embedded objects to share the same domain name as the container page. Origin-first CDN: two tcp connections are required.

29. 29 Request distribution in CDNs DNS/Balancing switch redirection. CDN DNS returns the IP addr of the balancing switch.

30. 30 Request distribution in CDNs Two-level DNS redirection This architecture distributes the DNS load among leaf DNS servers and allows client DNS servers to use nearby DNS servers for queries while they cache high-level DNS responses.

31. 31 Problems in DNS-based request distribution Originator problem: large networks often have a few DNS servers to handle clients’ requests. The client may be far away from the client DNS. Hidden load factor problem. Client DNS masking problem (recursive resolution of requests).

32. 32 Server Selection Metrics Proximity metrics Server load metrics Aggregate metrics Passive measurements obtain metrics by simply observing the normal operation of the system. Active measurements involve actions that the system performs only for the purpose of obtaining the metric. Synchronous/Asynchronous measurements.

33. 33 Proximity metrics Geographical distance Number of network routers Number of AS http://www.caida.org/home/

34. 34 Server Load Metrics Number of connections Number of requests Ready queue length Response time >> uptime 5:12pm up 30 day(s), 7:01, 3 users, load average: 0.11, 0.09, 0.09

35. 35 Aggregate metrics Tcp ping latency Icmp ping latency http request latency Download latency, download time