MARCH/APRIL 2005 Copublished by the IEEE CS and the AIP 1521-9615/05/$20.00 © 2005 IEEE 51PERSPECTIVESP ERSPECTIVES IN C OMPUTATIONAL S CIENCEand candor rarely encountered in a sin-gle work, the authors describe an evo-lutionary past drawn from their vastexperience and project an enticing andcompelling vision of HPC’s future. Yet, the underlying assumptions im-plicit in their treatment, particularlythose related to terminology and dom-inant trends, conflict with our own ex-perience, common practices, and sharedview of HPC’s future directions. Takenfrom our vantage points of the Top500list,2the Lawrence Berkeley NationalLaboratory NERSC computer center,Beowulf-class computing,3and researchin petaflops-scale computing architec-tures,4we offer an alternate perspectiveon several key issues in the form of aconstructive counterpoint.A New PathTerminology and taxonomies are sub-jective, with common usage dictatingpractical utility. Yet, in spite of its im-perfections, technical nomenclaturecan be a powerful tool for describing,distinguishing, and delineating amongrelated concepts, entities, and pro-cesses. Bell and Gray incorporate afundamental assumption throughouttheir reasoning, which, although de-fensible and advocated by notable re-searchers,5corrupts the terminology’spower as a tool to represent and differ-entiate. Specifically, their paper impliesthat essentially every parallel systememploying replicated resources is acluster. In this well-intentioned effortto provide a unifying principle,though, the authors have eliminated apowerful concept even as they in-tended to reinforce it. The concept ofthe commodity cluster has driven animportant trend in parallel processingover the past decade, delivering un-precedented performance-to-cost andproviding exceptional flexibility andtechnology tracking. By expanding thescope of the term, they’ve deprived thisimportant term of its seminal meaningand implication. One objective of this article is to re-store the strength and value of the term“cluster” by degeneralizing its applica-bility to a restricted subset of parallelcomputers. We’ll further consider thisclass in conjunction with its comple-menting terms constellation, Beowulf class,and massively parallel processing systems(MPPs), based on the classification usedby the Top500 list, which has trackedthe HPC field for more than a decade.As Bell and Gray convincingly artic-ulate, the impact of Moore’s law andthe economy of scale of mass-marketcomputing components in easily inte-grated ensembles will have a signifi-cant, even dominant, impact on theevolution of high-performance systemsin the near future. The Top500 list al-ready clearly reflects this trend withthe vast majority of all systems repre-sented on the list being products ofsome form of clustering. Moreover, asBell and Gray point out, Beowulf-classclusters are having a significant impacton medium-to-high-scale systemsthroughout the science and technicalcomputing arena as well as in the com-mercial sector. Also referred to asLinux clusters or PC clusters, Beowulfsare perhaps more widely used than anyother type of parallel computer be-cause of their low cost, flexibility, andaccessibility. Indeed, among the top 10systems on the most recent list (No-vember 2004 at www.top500.org), fiveare commodity clusters, three of whichare Linux clusters, not unlike the orig-inal Beowulf-class systems, and two areconstellations. One consequence of the progress an-ticipated beyond what Bell and Grayenvisioned is the form and content offuture computer centers, which willevolve as available technologies andsystem architecture classes advance. In-stead of becoming obsolete, the com-puter center’s role will likely grow inimportance, evolving to meet the chal-lenges of new architectures, program-ming models, mass storage, and acces-sibility via the Grid. The emergence ofBeowulf and other commodity clusterswill definitely alter the mix of resourcesthat will comprise a medium-to-large-sized computer center, but the respon-HIGH-PERFORMANCE COMPUTING:CLUSTERS, CONSTELLATIONS, MPPS,AND FUTURE DIRECTIONSBy Jack Dongarra, Thomas Sterling, Horst Simon, and Erich StrohmaierIN A RECENT PAPER,1GORDON BELL AND JIM GRAY PUT FORTHA VIEW OF THE PAST, PRESENT, AND FUTURE OF HIGH-PERFORMANCE COMPUTING (HPC) THAT IS BOTH INSIGHTFUL ANDTHOUGHT PROVOKING. IDENTIFYING KEY TRENDS WITH A GRACE52 COMPUTING IN SCIENCE & ENGINEERINGsibilities and services that necessitatebuilding such facilities will continue tobe critical, especially to the high-endcomputing and large data archive com-munities. Already we see in the US De-partment of Energy and the US Na-tional Science Foundation sector thedevelopment of new and larger com-puting centers to house the next gener-ation of high-end systems, includingvery large Beowulf clusters. The com-puter centers of the future will becharged with the administration, man-agement, and training associated withbringing these major resources to bearon mission-critical applications.This article, while congratulatingBell and Gray in opening up this line ofdiscourse, offers a constructive expan-sion on their original themes and seeksto correct specific areas of theirpremise with which we take exception.The long-term future of HPC archi-tectures will involve innovative struc-tures that support new paradigms ofexecution models, which in turn willgreatly enhance efficiency in terms ofperformance, cost, space, and powerwhile enabling scalability to tens orhundreds of petaflops. The conceptualframework offered here implies thedirections of such developments andresonates with recent advances beingpursued by the computer architectureresearch community.Commodity ClustersBell and Gray, in conjunction with theirdistinguished colleagues, see an impor-tant unifying principle emerging inHPC’s evolution: the integration ofhighly replicated components (many ofwhich were designed and fabricated formore general markets) as the drivingforce for a convergent architecture.They call this architecture a cluster anddistinguish it only from the minority setof vector supercomputers (such asNEC SX-6 and Cray X1) that exploitvectors in custom processor architec-ture designs. This convergent architec-ture model of the evolution of super-computer design is compelling, readilyapparent, and wrong. We respectfullyassert an alternate perspective that isrich in detail and has value in its abilityas an enabling framework for reasoningabout computing