The Optimal CPU and Interconnect for an HPC Cluster

The classic mainframe solutions in the HPC area consist of either SMP (symmetric multi-processor) or as MPP (massively parallel processing) systems. These are made up of incompatible limited-lot- production hardware devices that have been optimised to fulfil a special purpose by a single vendor. Those big-iron systems make use of shared memory architectures in which all processors share common main memory. In contrast to that, HPC (high performance computing) cluster solutions are made up of a number of commercially available computer systems using off-the-shelf hardware that is not restricted to a particular application. HPC clusters typically use distributed memory structures, that is a decentralised form of main memory scattered over individual nodes which is addressed by the CPU. The cluster is controlled by at least one administrative computer whose task is to make the system available to users with the help of a auxiliary program. From the viewpoint of the user a cluster consists of a software interface whose purpose is making resources available to user applications. This interface, also referred to as the middleware layer, runs on top of the operating system. Jobs are assigned to processors over a dedicated network. In what follows we will take up aspects of High Performance Compute Cluster that are relevant for procurement planning.