Load Balancer

A hardware appliance or software program that distributes incoming client requests across a set of configured servers, providing fault tolerance and improving throughput.

Layer 4 Switch

A basic load balancer that makes forwarding decisions on IP address and TCP/UDP header values, working at the Transport layer.

Layer 7 Switch

A more modern load balancer that makes forwarding decisions based on application-level data, like URLs or data types like audio/video streaming.

Clustering

A load balancing technique where a group of servers are configured as a unit and work together to provide network services. If one of the multiple redundant processing nodes go down, connections can failover to a parallel, working node. However, to clients, the node cluster looks like a single server. This is done using virtual IP addresses. While each node is configured with a private connection, identifiable using it's real IP address, a redundancy protocol (e.g. CARP) ensures that an 'active' node 'owns' the virtual address and responds to incoming connections. The protocol also has some sort of a heartbeat mechanism to failover to a parallel passive node if the active one goes down.

There are two main types of node clustering, active-active and active-passive.

Active-Active Clustering

Say you have two nodes clustered to serve a webpage. In this case, both nodes are accepting connections and processing requests concurrently, allowing you to efficiently use all the hardware resources. However, should one node go down, all the traffic would failover to the other one, greatly reducing performance.

Active-Passive Clustering

The inverse of active-active clustering. In this case, there is only one active node that is listening for and processing incoming requests. The big benefit of this is that performance doesn't immediately die if a node goes down. However, the maintenance cost might be higher relative to the amount of performance you're getting since only one node is active at a time.

In standard active-passive configurations, each active node needs an accompanying passive node. However, there are other configurations (N+1 and N+M) that allow for fewer passive nodes than active nodes to reduce costs.