Hardware Components: Active and Passive
What are the major components in a Linux cluster? I tend to divide the collections of cluster components into hardware and software classes. We'll discuss software components in the next section. Within the hardware grouping are both active and passive elements. The active components, such as switches, compute slices, file servers, storage devices, SAN switches, and so on are the "glamour kids" on the block—we pay a lot of attention to them because they're smart and fun to work with (sometimes).
The passive components—cables, racks, system-mounting rails, power distribution units (PDUs)—are the "poor cousins" of their active relatives, at least from the glamour standpoint. Surprisingly, properly chosen passive components, in their supporting roles, are extremely important to the overall success of the cluster. You overlook the importance of cables and a cable-documentation system at your peril. Cheap racks that have poor airflow, torque out of shape, or cannot be shipped with equipment in them are a false economy.
Figure 1 shows a sample cluster's physical layout of 18-CPU cluster and associated hardware components. There are nine two-CPU compute slices, one master node with fibre channel-attached storage, 18 SCSI disk-storage trays (two per compute slice) with four disks each providing per-system local scratch space, a core Ethernet switch, and a Myrinet high-speed interconnect switch. This cluster is similar to one I built for a customer who was doing reservoir simulation in the oil and gas industry. Notice the passive and the active components in the example.
Figure 1 Example of a cluster's physical rack layout.