DNS-based approach

In this approach, server side authorized DNS maps domain name to IP address of one of the nodes of the cluster, based on various scheduling policies. Selection of replica occurs at server side DNS so it does not suffer from applicability problem of client-side mechanisms. But DNS has limited control over requests reaching at server because of caching of IP address mapping at several levels viz., by client softwares, local DNS resolvers, intermediate name servers, etc. Besides the mapping, a validity period for this URL to IP address mapping, known as Time-To-Live (TTL) is also supplied. After expiration of TTL period this mapping request is again forwarded to authorized DNS. Setting this value to very small or zero does not work because of existence of non cooperative intermediate name servers and client level caching. Also, it increases network traffic and DNS itself can become bottleneck.
Several DNS based approaches are discussed in [9]and [12]. DNS based algorithms can be classified on the basis of the scheduling algorithms used for server selection and TTL values.

• Constant TTL algorithms : These are classified on the basis of the system state information used by DNS for server selection. The system state information can include both client and server state information, like load, location etc.
  1. System stateless algorithms : Most simple and first used algorithm of this type is round robin (DNS-RR). It was used by NCSA (National Center for Supercomputing Applications) [24] to handle large traffic volume using multiple servers. In this approach, primary DNS returns IP addresses of servers in the round robin fashion.

     It suffers from uneven load distribution and server overloading, since large number of client from same domain (using same proxy/gateway) are assigned same server. Also, whole document tree must be replicated on every server or network file system should be used.

  2. Server state based algorithms : A simple feedback mechanism from servers about their loads is very effective in avoiding server overloading and not giving IP address of unavailable servers. The scheduling policy might be to select the least loaded server any time.

     This approach solves overloading problem to some extent yet control over requests is not good because of caching of IP addresses. Some implementations try to solve this problem by reducing TTL value to zero but it is not generally applicable and puts more load on DNS.

  3. Client state based algorithms : In this approach, two types of information about clients, the typical load arriving to system from each connected domain (from same proxy/gateway) and the geographical proximity can be used by DNS for scheduling.
     Requests arriving from domains having higher request rate per TTL value can be assigned to more capable server. Proximity information can be used to select nearest server to minimize network traffic.

     One mode of Cisco Distributed Director [11] takes client location (approximated from client's IP address) and client-server link latency into account to select the server by acting as primary DNS.
     This approach also suffers form same problem experienced by Server state based algorithms.

  4. Server and Client state based algorithms : Cisco Distributed Director takes server availability information along with client proximity information into account while making server selection decision. These algorithms can also use various other state estimates for server selection. Such algorithms give the best results.

• Dynamic TTL algorithms : These algorithms also change TTL values while mapping host name to address. These are of two types [12]:
  1. Variable TTL algorithms : As server load increases these algorithms try to increase DNS control over request distribution by decreasing TTL values.
  2. Adaptive TTL algorithms : These algorithms take into account the domain request rate (number of requests from a domain in TTL time period) and server capacities, for assigning TTL values. So a large TTL value can be assigned for a more capable server and less TTL value for those mappings that have high domain request rate.
     These are most robust and effective in load balancing even in presence of skewed loads and non-cooperative name servers, but these don't take geographical information into account.

DNS based approaches are more suitable for static replication schemes and are less suitable for dynamic replication schemes because changing place of replicated object may require change in mapping. In general these approaches suffer from limited control over request problem due to caching of resolved IP addresses at various levels.