Performance Characterization of the World's Most Powerful Supercomputers

Over the last decade, the architectural spectrum of large-scale parallel systems has been dominated by the cluster-of-workstation paradigm. These systems are characterized by loose hardware integration between the networking and the CPU/memory components, and lack the efficiency of custom-designed computing platforms. As a result, while peak processor performance increases exponentially according to Moore's Law, improvements in sustained performance lag far behind. The problem is expected to exacerbate by the end of this decade, because traditional uniprocessor scaling trends are slowing down as most of the benefits of instruction level parallelism and pipelining have been mined, and clock frequency increases are limited by power dissipation concerns. In the same time frame, mission-critical applications will have computational requirements that are at least two orders of magnitude larger than current levels. Numerous recent developments in high-end computing (HEC) system design have the potential to overcome these deficiencies, via the use of parallel vector-based platforms; ultra-scale systems built from low-power components; scalable shared-address space architectures; and commodity clusters utilizing custom interconnects. To understand the tradeoffs between the myriad of architectural design choices, comparative application performance data must be readily available to the HEC community at large, as performance of full-scale scientific applications is the final arbiter of supercomputing platforms. However, evaluating large-scale scientific applications using realistic problem sizes on leading HEC platforms is an extremely complex process, requiring coordination, communication, and management of a team of application scientists coming from highly disparate backgrounds. A detailed application analysis reveals algorithmic limitations to scalability and indicates where a particular code is poorly suited to certain architectural features. Direct comparison of applications across architectures, combined with a careful analysis of the problem and implementation characteristics, provides invaluable insight into the suitability of a given machine type for a particular class of scientific methods. This special issue is a collection of papers that present and analyze performance results from the latest generation of HEC platforms, representing the world's most powerful supercomputers. Several of these papers were presented at the 2006 SIAM Conference on Parallel Processing for Scientific Computing held in San Francisco. The first paper by Oliker et al. compare and analyze the performance of four application codes drawn from the areas of magnetic fusion, plasma physics, astrophysics, and materials science on leading vector and superscalar architectures. The authors also present several major implementation innovations, and conclude that modern parallel vector systems have tremendous potential to attain impressive …