In a rapidly growing and interdisciplinary field like information technology, terminology has always been a concern, as a given word may denote different ideas across groups and across time. A term is meaningful only if it helps to define a useful class of things, and is neither too broad, nor too narrow. Debate over the scope of the term “cluster” forms the main thread in this paper, which is part of the “Perspectives in Computational Science” column.
The paper, expressing disagreement with the definition of “cluster” in Bell and Gray’s article [1], argues for a narrower scope to drive home select important characteristics. This class of clusters, defined as an integrated collection of nodes that are capable of standalone operation, is expected to play an important role in the continuing evolution of parallel computing toward petaflop computing. Other architectural developments are contrasted against the role of clusters in this regard. The paper also visualizes the traditional notion of a computer center adapting itself to manage cluster environments.
Extending the argument on the definition of a cluster, the paper outlines a new framework for classifying high-performance computing systems. Based on current trends and issues, four parameters are identified as the basis for this classification: clustering, namespace, parallelism, and management of latency and locality. An initial set of options is provided for each, to be refined further by the community.
The arguments put forth by the authors are interesting, and worth a closer look. This may help inspire a cleaner redefinition of common terms in the area of parallel computing, as well as provide a useful way to characterize and compare parallel and distributed computing systems.