By Rich Farrell
Network World, 9/22/97

There's more than one way to balance traffic load across multiple Web servers. Rotary Domain Name System (DNS) is an older technique that these devices replace. If you're using rotary DNS for load balancing, consider the devices reviewed here for improved performance and fault tolerance.

Rotary DNS is a way of modifying DNS entries so entries for a virtual name can service requests for multiple machines. Machines move to the head of the list one by one in a round-robin fashion. If there are three machines listed for a given virtual name, the order of the three will be switched periodically so some of the requests will see one machine as the preferred translation for a name, other requests will see the second, and still others will see the third.

Because standard DNS is used by rotary DNS, it is not possible to predict and control client caching, nor, therefore, the actual load on the servers. Adding and removing machines can take many hours as DNS changes are propagated through the network, and a lot of traffic could be sent to a machine that has been removed for hours or even days. Rotary DNS changes the IP address associated with a virtual site - it does not have any effect on which protocols or ports can be used with that address.

Employing dedicated load-balancing hardware offers many advantages over rotary DNS. Because the load-balancing devices map incoming requests to real servers and traffic passes through them to and from the real servers, a count of open connections can be established. Round-robin load balancing treats all servers as equal, similar to rotary DNS but without propagation delays or caching issues. Web farms with several devices all configured in a similar fashion and all of the same capacity will find round-robin to be a predictable, efficient algorithm.

Least connections is useful when servers in the Web farm are not of the same capacities (smaller/slower vs. larger/faster machines) or are running different software. Connections take longer on slower machines, so as the number of outstanding connections increases, traffic migrates to machines with fewer connections.

Because not all real machines in a Web farm are necessarily equal in capacity or in the tasks they can perform, system managers may need the ability to assign "weights'' to each machine, giving the load-balancing device hints as to how to balance the traffic. Weighted percentage lets you affect how the servers are assigned proportions of traffic by percentage. This gives the system administrator the ability to favor one machine over the other rather than relying completely on the load-balancing software to figure out the inequities of the various machines in the server farm.

The load-balancing algorithm known as fastest keeps track of the length of time a machine takes to fulfill a request and refers traffic to the server with the fastest response rate. Maximum connections can be used to ensure machines don't go above thresholds that are deemed to be the maximum desired.

With server daemon monitoring, each real machine runs a daemon process that the load-balancing machine pools to ascertain the relative health and capacity of the real machine. This algorithm only works with Unix and Windows NT Server operating systems.