Teraflops computing: A challenge to parallel numerics?

Abstract

Following in the wake of the Accelerated Strategic Computing Initiative (ASCI) of the US Department of Energy, in the forthcoming years powerful new supercomputers will be brought into the market by the manufacturers participating in the high-performance computing race. Hence, the large-scale computing facilities in the key research centers and industrial plants world-wide will surpass the teraflops performance barrier, too. The parallel architectures will be further extended to hierarchically clustered parallel computers mainly based on commoditychip processors and SMP nodes tying together possibly tens of thousands of processing elements. In addition, heterogeneous computing and metacomputing will determine future large-scale computing by interconnecting supercomputers of diverse architectures as giant supercomputer complexes. These developments will challenge not only system reliability, availability and serviceability to novel levels, but also interactivity of concurrent algorithms and, in particular, adaptivity, accuracy and stability of parallel numerical methods.