On the Value of Diversity: An Insider's View of High-Performance Computing

KFA is one of the largest big-science research centers in Europe. Its scientific and engineering activities are ranging from fundamental research to applied science technology. KFA's Central Institute for Applied Mathematics (ZAM) is running the large-scale computing facilities, central servers, and network systems at KFA. ZAM is also providing supercomputer capacity to the scientific community all over Germany; this broad service has been a novelty in 1987 when a new user meta-structure named HLRZ (Höchstleistungsrechenzentrum) has been established in order to further promote computational science and engineering in Germany. In 1996, ZAM's facilities are upgraded, in addition to a 16-node IBM SP2 and a 140-node Intel Paragon, by a heterogeneous new supercomputer complex consisting of CRAY J90 (20 CPUs) T90 (12 CPUs), and T3E (512 processors). Based on the recent recommendations of the German Science Council, also several large university computing centers are preparing the necessary steps to reenforce the apex of the structural pyramid of modern scientific computing.While the successes of the past two decades in scientific supercomputing have been achieved essentially by the technical breakthrough of the vector-supercomputers, the future of supercomputing is focussed on massively parallel computers. Besides system architecture, node performance, and interconnection topology, important issues are genuine parallel algorithms, programming tools, and programming models. Message Passing on parallel computers with distributed memory is the only efficient programming paradigm available today; from a user's point of view, however, rather the concept of shared virtual memory will be capable to serve as the effective basis in order to bring computing on massively parallel systems from just a computer science toy to the technological breakthough.Since large applications, in general, are algorithmically heterogeneous, they will probably benefit best from the diversity of computer structures if these are cooperating and combined via high speed links to heterogeneous supercomputing ensembles