Preconfigured Cluster Hardware
If you don't want to go to the effort of racking and cabling your cluster's hardware, other options can significantly reduce the design and purchase time. First, most major hardware vendors have preconfigured hardware offerings similar to Hewlett-Packard's "cluster blocks" or IBM's eServer cluster. These are cluster hardware configurations comprising compute slices, network switches, high-speed interconnect (if desired), console management, and head nodes—all pre-racked and pre-cabled. Another viable option is to go with high-performance computing (HPC) vendors such as Linux Networx that offer ready-to-go cluster configurations. Unless you have lots of free time, or employees who are willing to work for pizza and beer, the preconfigured route may be your best option.
A cluster for HPC may be very complex and require a lot of infrastructure to support it. Some of the context for such a cluster is shown in Figure 2, which represents a mixture of logical and physical cluster components. A cluster of this size and complexity is not a recommended project for a first-time cluster builder. It's better to start small, particularly if your organization has no cluster experience and a low tolerance for project overruns.
Figure 2 Supporting infrastructure for a clustered system.
Some of the physical elements we're discussing are tied to the overall software architecture of the cluster to be designed or built. For example, the high-speed interconnect (HSI) and other networks help tie the cluster together, and so they exert a major influence on the overall software architecture. The architectural components of our cluster are shown in Figure 3, with the hardware-related pieces outlined in dashed boxes. We'll consider the software components next.
Figure 3 Cluster architectural components—hardware elements.