Beowulf-class computing systems are part of a large family of parallel computer architectures. Indeed, different hierarchical taxonomies of parallel computers are possible depending on the primary issues chosen to distinguish among them. Here, degree-of-coupling is applied loosely to reflect a combination of bisection bandwidth and worst case communication latency. At the top level of this taxonomy from most loosely coupled to most tightly coupled are meta-computers, clustered computing, multiprocessors, and vector based systems. Meta-computers harness widely separated computer systems by means of wide area networks with latencies, i.e., the time lapse between sending and receiving a message, measured in the tens of milliseconds. Clusters are loosely coupled collections of stand-alone computing elements, usually in the same location and administrative domain, with latencies between 10 to 100 microseconds. MPPs comprise many processors tightly connected within the same unit exhibiting latencies of between half a microsecond and 5 microseconds. Vector computers pipeline computations so that the effective latency between successive operand sets can be measured in the tens of nanoseconds.
Of these systems, clusters, with a history that goes back to almost the dawn of digital electronic computing, are directly relevant to Beowulf. Clusters comprise essentially independent and self supporting systems that are networked to provide a means of interaction. Clusters differ from meta-computing in that they are relatively local, under the same administrative domain, and incorporate a dedicated network. Clusters can be implemented with networks of servers, workstations, or PCs. Another form of clustering is the use of an existing local area network of desktop computers which can be used collectively when the computers are not serving their respective users. Known alternatively as "cycle harvesting" or "workstation farms", this form of cluster can provide a valuable resource under certain circumstances. Within the family of clusters, Beowulfs fall under clusters of PCs, also referred to as ''piles of PCs."
PC clusters are subdivided into additional groups, one of which is Beowulf-class systems. Another group is NT-clusters using PCs but running the Microsoft Windows NT operating system. Yet another type of PC cluster employs specialty networks such as SCI or that from the Princeton SHRIMP project, providing hardware support for distributed shared memory. Although identified as a single category of PC clusters, Beowulf systems themselves exhibit wide diversity in configuration and operation.
