Iptables gives admins the ability to set up clusters and distribute the load. But what about failover?

Load sharing technologies often rely on a central system or application that distributes the work evenly over the members of the cluster. The Linux Virtual Server project implements this on Linux. To avoid a single point of failure, the central instances should be highly available and continuously monitored by a routine that checks the systems and responds to errors or lost signals. If you prefer to avoid a central load sharing instance entirely, the iptables CLUSTERIP target is an alternative. CLUSTERIP is a simple and inexpensive technique for load sharing that is already part of the Netfilter code, and although this feature is not entirely stable, the technology is quite impressive.
In CLUSTERIP, the cluster nodes share a common address, and each node uses a hash algorithm to decide whether it is responsible for a connection. Admins can assign responsibilities to a node via /proc/net/ipt_CLUSTERIP, influencing load sharing, or switching interactively or by means of dynamic scripting. Stonesoft products have had this functionality for a while, and it works well.
Iptables clusters do not have a built-in heartbeat mechanism to check the health state of the nodes, remove broken systems from the cluster, or tell other nodes to take over the load of the failed system. Many failures are heralded by tell-tale signs, however, that give the ailing node the ability to voluntarily leave the cluster in good time. In this article, I show the possibilities of combining the CLUSTERIP target of iptables with a script controlling the cluster.

Read full article as PDF »

Iptables gives admins the ability to set up clusters and distribute the load. But what about failover?