The Internet naming server provides support for load balancing by having a one-to-many mapping from hostnames to IP addresses [7]. Early on, a round-robin DNS was used with success for the NCSA Web server [15]. Replicating services using naming mechanisms has a number of disadvantages. First, substantial caching happens in the name resolution hierarchy of DNS. This significantly adds inertia before the load balancing effect of round-robin resolution kicks in. To make DNS based schemes react more quickly, caching must be limited. This in turn puts substantial load on the name servers. Second, the replication happens at hostname level, which makes the selective replication of services awkward. At the least, services that are replicated differently from each other would require different host names. Third, transparent reconfiguration after a host failure would not be possible. Clients typically do not resort to repeated name resolution when a server is not reachable. DNS based approaches to recover from host failures work rather well for the Web because URLs mostly contain host names and HTTP requests are largely idempotent.