September 2008 Josilene Aires Moreira

 Overview  CDN Topology  CDNs nowadays  Contructing a CDN ◦ Basic model ◦ Modules ◦ Characteristics  References

 Origin Servers  Surrogate Servers (mirrow servers)

 First evolved in 1998, replicate content over many web servers to deal with flash crowds  Geographically distributed web servers to improve performance and scalability  Improves network performance through content replication ◦ Maximizing bandwidth ◦ Improving accessbility ◦ Mantaining correctness  Moves content to edge servers located close to the users

 Host thirdy-part content  Any digital content  Static ◦ Html pages ◦ Images ◦ Documents ◦ Software  Streaming media ◦ Audio ◦ Video  Content services ◦ E-commerce service Source of the content Large enterprises Web service providers Media companies New broadcasters

 Origin Server  Surrogate Servers ◦ A set of replica servers that cache the origin’s server content  Routers and network elements ◦ Deliver content requests to the optimal location and the optimal surrogate server  Accounting mechanism ◦ Provides logs and information to the origin server

 Advantages ◦ Reducing the customer need to invest in Website infrastructure ◦ Decreasing the operational effort and costs of managing such infrastructure ◦ Bypassing traffic jams  Data is closer to the user  There is no needs to transverse all congested pipes ◦ Improving content delivery quality, speed and reliability ◦ Reducing the load on origin servers

 Akamai Technologies ( is the market leader ◦ (80% of the overall CDN market) in providing content delivery services. It owns more than 12,000 servers over 1,000 networks in 62 countries.  Mirror Image Internet, Inc. ( ◦ supports surrogate servers located in 22 cities around the world (North America, Europe, and Asia), which provide a range of value-added services, from content distribution to media streaming and managed caching.  Inktomi, a Yahoo Company ( ◦ Provides managed services for global load balancing, failover, content delivery, and streaming media using more than 1,000 surrogate servers worldwide.  LimeLight Network ( ◦ provides a suite of services (including music download and subscription services, video game developers and distributors, movie/video download services, and so forth) and supports surrogate servers located in 72 locations around the world (Asia, the U.S., and Europe).

 Codeen ◦ Codeen is developed at Princeton University, USA. It provides caching of Web content and redirection of HTTP requests. ◦ It is an academic testbed content distribution network built on top of PlanetLab. This testbed consists of a network of high performance proxy servers. ◦ Each of these proxy servers acts both as request redirectors and surrogate servers. They provide fast and robust Web content delivery service through cooperation and collaboration.  Coral CDN ◦ Coral is a free, peer-to-peer content distribution network designed to mirror web content. ◦ Coral is designed to use the bandwidth of volunteers to avoid congestion and to reduce the load on websites and other web content providers in general.  Globule ◦ Globule is an open-source collaborative content delivery network developed at the Vrije Universiteit in Amsterdam. Globule aims to allow Web content providers to organize together and operate their own world-wide hosting platform. It provides replication of content, monitoring of servers and redirecting of client requests to available replicas.

 Origin  Surrogate  Content manager  Redirector  Clients  Access network On Content Delivery Network Implementation, Molina Moreno et al, 2006

 Replica servers ◦ Acts as proxy/cache servers ◦ Store and delivery content  Components ◦ Portal = http-based web-server  Provides access to the content stored in the CDN ◦ Streaming server  A media streaming server to distribute multimedia content ◦ Surrogate Database (DB)  Contains a list of all available streaming sessions, the objects stored in the surrogate and information for the management of the CDN  Research Questions ◦ Number of Surrogate Servers ◦ Surrogate Placement Algorithms (Greedy, Hot-spot, Tree-based,…)

 Research Questions ◦ Number of Surrogate Servers ◦ Surrogate Placement Algorithms  Placement Algorithms  Content distribution includes the placement of surrogate servers to some strategic positions, close to the clients.  Some theoretical approaches model the problem as the “center placement problem”: for the placement of a given number of centers, minimize the maximum distance between a node and the nearest center. ◦ Greedy ◦ Hot spot ◦ Tree-based

 Control the media objects stored in each Surrogate  Provide this information to the Redirector, to get each client served by the most suitable Surrogate or Peer  Content Locator Determines ◦ Number of replicas of a media object ◦ In which surrogate servers a new object has to be stored ◦ Elimination of non-popular objects from the surrogates ◦ Interaction of the CDN within the origin servers ◦ Update media objects in the surrogates ◦ Move media objects among surrogates  DB Content ◦ Stores the information managed by the Content Locator

 Research Questions  Content Outsourcing  Related to the relationship between surrogates ◦ Cooperative push-based  The content is initially pushed from Origin Server  The surrogate servers cooperate to reduce replication  This approach has not yet been adopted by a CDN provider ◦ Uncooperative pull-based  Clients’ requests are directed to their closest Surrogate Server  If there is a cache miss, the request is redirected to another Surrogate or to the Origin Server  Akamai and Mirror Image use this approach ◦ Cooperative pull-based  Client requests are directed to their closest surrogate  In case of cache misses, the surrogate servers use a distributed index to find nearby copies of the requested objects  Coral (academic CDN) has implemented this approach

 Research Questions (cont.)  Content Selection Related to how the content is replicated at the surrogates ◦ Entire replication  High disk prices  Web objects are increasing  Updating is a problem ◦ Partial site content selection and delivery  Empirical  Uses heuristics to decide the content to be replicated at the edge servers (uncertainty in choosing right heuristics)  Popularity-based  Most popular objects are replicated  Object-based  Content is replicated in units of objects that gives the maximum performance gain (high complexity to implement on real applications)  Cluster-based  Content is grouped based on correlation or access frequency and replicated in units of clusters (high complexity)

 Research Questions (cont.)  Caching Techniques Related to how the caches manage a cache miss ◦ Intra-cluster caching  Query-based  On a cache miss, a CDN server broadcasts a query to others cooperating CDN servers (significant query traffic)  Digest-based  Each CDN server has a map of content held by the other cooperating surrogates (update traffic overhead)  Directory-based  A centralized server keeps information of all the cooperating surrogates in a cluster (potential bottleneck = single point of failure)  Hashing-based / Semi-hashing based for streaming  Cooperating CDN servers maintain the same hash function  More efficient, highest content sharing efficiency

 Research Questions (cont.)  Cache update ◦ Periodic Update  Most common method ◦ Update propagation  An updated version is delivered to all caches whenever a change is made into the origin server ◦ On-demand update  The latest copy of a document is propagated to the surrogate cache server based on prior request for that content. The content is not updated until be requested ◦ Invalidation approach  An invalidation message is sent to all surrogate caches when there is a change at the origin server, and the caches need to fetch an updated version later

 Provides intelligence to the system ◦ Estimates the most adequate surrogate server for each different client request  CDN DNS ◦ Accept requests from the client local DNS and route the client to the most adequate surrogate (several approaches, Akamai uses servers in two hierarchies) ◦ DB DNS  Stores information about addresses and names of the surrogate servers, and some additional information

Redirector (Request-routing)  Monitor Module (Main knowledge gatherer of the system) ◦ Gets statistical information from different key elements of the CDN architecture ◦ Conducts a variety of measurements to obtain information about the network and the CDN components ◦ Uses SNMP to get data from the surrogates to estimate the RTT between Clients and Surrogates ◦ DB Monitor : Stores all information  Redirection Algorithm ◦ Selects the optimal surrogate server on the basis of the information gathered by the Monitor

 Research Questions  Which one is the appropriate surrogate server? ◦ The replica server ‘closest’ to the client  Decision based on Metrics (‘closest’) ◦ Network proximity ◦ Client perceived latency ◦ Distance ◦ Replica server load  Request Routing Algorithm ◦ Non-adaptive  Round robin, prediction on load of servers (number of requests), client-server distance ◦ Adaptive  Network proximity (path length), Client perceived latency  Akamai uses a complex combination of server load, reliability of paths, bandwidth available to a replica server  Cisco DistributedDirector uses inter-AS distance, intra-AS distance, end-to-end latency

 Pallis, G. and Vakali, A., “Insight and Perspectives for Content Delivery Networks”, Communications of the ACM, vol. 49, no. 1, ACM Press, NY, USA, January  Moreno, B. M.; Salvador, C. P.; Domingo, M. E.; Peña, I. A. & Extremera, V. R. (2006), 'On content delivery network implementation', Computer communications 29(12),  Pathan, A. K. & Buyya, R. (2007), 'A Taxonomy and Survey of Content Delivery Networks', Technical report, The University of Melbourne, Australia