Server-based approach

This approach allows two-level dispatching, first by cluster DNS and later each server may reassign a received request to one of the other server in the cluster. This solves the problem of non-uniform load distribution of client request and limited control of DNS.

• HTTP redirection by Server
  The approach is used in SWEB [3]. First request reaches to host in cluster using normal DNS resolution but it can further redirect request to other server. It does second level dispatching through the redirection mechanism of the HTTP protocol. This redirection may depend on the load of server or may be done in a round robin fashion. The servers need to exchange status information periodically for taking redirection decisions but this cost is negligible with respect to traffic generated by client requests. Its main disadvantage is duplication of TCP connections and hence increased delay in response.

• Packet Forwarding by Server
  In this approach, first level scheduling is done using round robin DNS mechanism, the second level dispatching is done by packet rewriting mechanism that is transparent to users. So first request reaches to any node in cluster, if that node figures out that other node is better for serving this request, node uses MAC address to reroute the packet to selected sever.

  It does not require HTTP request redirection hence it is better in terms of latency time. The server selection can be stateless i.e. based on hash function or based on load information on servers. If loading information is used for rerouting, server need to exchange load information among themselves. Also this scheme can work with both LAN and WAN based solution.

• Akamai's Approach
  Akamai's approach [2] is very different. In their approach, URLs of objects embedded in HTML page, like images, Java Applets, multimedia components etc., are modified by proprietary software Launcher running at server, to the URLs of the objects available at any Akamai server nearest to client. It is claimed that these embedded objects comprise nearly 70% of typical page in overall bytes. A map of current Internet traffic conditions, the loads of all Akamai servers worldwide, and the locations of Internet users is built for selection of server. This map is updated once per second. While making selection of server, it is made sure that no server is overloaded and number of servers containing replica is proportional to number of requests for the object.
  This approach is very useful when page contains large multimedia objects. It requires protocol for getting information about other servers distributed geographically, and client location. It scales geographically well but it also requires pages to be modified according to the client location.