We used a HP Procurve switch to interconnect the nodes and our disk server. Each node uses dhcp from our disk server for network configuration. We linked our current cluster to our old cluster using a single crossover cable.
We decided that a standard batch system such as the Portable Batch System was inappropriate for our computer usage patterns. Batch systems are effective in insuring that CPUs are in constant use but they have several limitations that make them difficult to use at SEP. The most significant issue is that SEP problems generally take up significant disk space and can have high IO needs. With a batch system you have little to no control over which node your process starts on. As a result you can pay a high IO cost. The other problem with a batch system is that its design parameters don't address the type of problems we encounter in a research environment. At best, a batch system determines the number of processes to use at start up. We occasionally have jobs that will take up to a week to run. A traditional batch system will either fill all available nodes with the job, making the machine unavailable for everyone else for a week or start it on such a few number of nodes that a week's job will take two months. Neither option is optimal.
To monitor the nodes we wrote a simple service that returns the result of the ps command when a certain port is accessed. This methodology proved more reliable than a simple rsh approach. From our entry point we periodically attempt to contact each machine and store load and usage information. For easier evaluation of this information we wrote a simple web interface. Currently we limit computer usage by assigning separate accounts, but make the usage transparent by using the same uid for both SEP and cluster accounts.
For console access we debated whether to use a Rocket Port Serial Hub or a series of switches. Both solutions introduced significant cabling and additional costs. For now we are using the Virtual Network Computing (VNC) developed by ATT for general console displays and attaching a monitor and keyboard when we experience problems with a node. So far this solution has proven effective but may not be optimal for a large system.