Ü 33 57

TISTESüflOi'i ST^.TEMEITT^A Ä„„,..r~~~A -f.-i-r rjuWc release; Distribution Unlimited Two years ago, researchers at Sandia National Laboratories showed that a massively parallel computer with 1024 processors could solve scientific problems more than 1000 times faster than a single processor. Since then, interest in massively parallel processing has increased dramatically. This review paper discusses some of the applications of this emerging technology to important problems at Sandia. Particular attention is given here to the impact of massively parallel systems on applications related to national defense. New concepts in heterogenous programming and load balancing for MIMD computers are drastically increasing synthetic aperture radar (SAR) and SDI modeling capabilities. Also, researchers are showing that the current generation of massively parallel MIMD and SM) computers are highly competitive with a CRAY on hydrodynamic and structural mechanics codes that are optimized for vector processors. INTRODUCTION Two years ago, Sandia National Laboratories won the Gordon Bell Prize, the Karp Challenge and a R&D 100 award for its pioneering research in massively parallel computing. A major contribution of this work was to show that speedups of greater than 1000 could be achieved on scientific problems using a first generation massively parallel machine with 1024 processors previously, conventional wisdom was that it would be difficult to use as many as one hundred processors to solve a single problem Researchers at Sandia and elsewhere are now showing that the current (second) generation of massively parallel computers manufactured by nCUBE, Thinking Machines, INTEL, and others can outperform traditional vector supercomputers on a wide range of applications." In the next few years, significant advances in massively parallel systems are anticipated as more processors are added and the computational power of each processor increases. Conventional vector processors, on the other hand, are not expected to make major gains in the near future. During the past decade, clockspeeds on the fastest supercomputers have increased by a factor of two. While a factor of two might be a major improvement in many fields, it is very small in the computer industry. By contrast, the microprocessor revolution has led to order of magnitude improvements in desktop computers. The slow increase of supercomputer speeds is attributable to a fundamental physical barrier namely, the speed of light Clockspeeds on vector supercomputers are now measured in nanoseconds. For reference, one nanosecond is approximately the time it takes light to travel a foot Dramatic increases in performance will not occur without a major technological breakthrough in, for example, high temperature superconductivity, that allows the size of these supercomputers to be decreased dramatically. It is becoming increasingly evident that massively parallel systems represent the future of supercomputing. All of the major supercomputer vendors, including CRAY, DEC and IBM, have recently announced major projects to develop massively parallel machines. The Japanese government is now investing heavUy in this technology and on a recent scientific visit to the Soviet Union, Sandia researchers learned that the Soviets have developed a 128 node parallel computer that gives CRAY-class performance. During the last two years, Sandia's massively parallel computing effort has grown to more than fifty researchers. Large-scale scientific calculations are being conducted on an nCUBE 2, a Connection Machine (CM-2), a first generation nCUBE/ten (sometimes referred to as an nCUBE 1), and a network of workstations. Sandia is also a member of a conEf 1 ill u V.5. CSIWBUTIONOFWISDO ^ sortium headed by Caltech that is purchasing a large INTEL Delta prototype. The nCUBE 2 at Sandia has 1024 processors, 4 Gigabytes of memory, and will soon have a 16 Gigabyte parallel disk system. Typical applications are executed at 1 to 2 billion calculations per second (Gigaflops) and the speed can be increased to over 10 Gigaflops by going to a system with the maximum possible 8192 processors. The CM-2 at Sandia has 16384 processors, 2 Gigabytes of memory and a 10 Gigabyte disk farm Currently, as much as half of the total usage of our nCUBE 2 and CM-2 is by external researchers at universities, other national laboratories and private companies. These massively parallel computers are being used at Sandia to solve a wide range of problems involving, for example, radar imaging, SDI tracking and correlating, shock physics, structural mechanics, fluid mechanics, heat transfer, the motion of charged particle beams through accelerators, molecular dynamics and quantum chemistry. In order to support these activities, research is conducted in mathematical algorithms, load balancing, domain decomposition, grid generation, and graphics. The work in mathematical algorithms includes both the development of new methods (e.g., parallel time stepping, which is a technique to accelerate time marching on parallel computers) and the parallel implementation of existing techniques (e.g. multigrid and conjugate gradient based schemes for solving systems of linear equations arising from the solution of partial differential equations)." Novel dynamic load balancing techniques have been constructed for Monte Carlo problems and particle simulations similar techniques are being investigated for hydrodynamics and shock physics computer codes. Work in graphics is being stressed because the problems being solved on massively parallel computers are large. One Gigabyte of data is often considered small in the applications of interest. It is next to impossible to analyze and evaluate results from such calculations without visualization. All of the large Sandia parallel machines (the nCUBE 2, CM2 and nCUBE/ten) have graphics systems and a Stardent GS2000 with a Application Visualization System (AVS) is available. This review paper discusses some of the codes that are being implemented on massively parallel computers at Sandia. Particular attention is given to codes that are used extensively to support national defense programs. The following sections discuss the parallel implementation and performance of radar imaging, SDI tracking and correlating, shock physics and structural mechanics models. SYNTHETIC APERATURE RADAR SIMULATIONS Synthetic-aperture radar (SAR) is often used in reconnaissance and remote sensing applications. SAR systems transmit and receive coherent broadband signals that are processed and focused to produce terrain images. These systems can be used at day or night and in all types of weather. Computer simulations are needed to assess the performance of SARs. These models use ray tracing techniques similar to those employed in computer graphics. In military applications, difficulties arise because objects of interest have many edges and multiple reflections can occur. Consequently, as many as a million rays are needed to provide adequate resolution. It can easily take more than an hour to construct a complicated image using a conventional supercomputer. This is due to low efficiencies resulting from a poor match between the ray tracing problem and vector architectures. Because thousands of orientations are often required, many problems of interest are intractable using CRAY-type computers. A massively parallel, MIMD version of a radar imaging code, SRIM, has been developed at Sandia. The beam is modeled as a collection of rays that leave the radar and bounce off targets, and a radar or optical image is constructed from ray-target intersections. The geometry of the objects and their reflection characteristics are specified a priori in the model (see Fig. 1). The output of SRIM is an idealized, noise-free radar image whose resolution is limited only by the number of rays used. A real radar image has noise and finite resolution, both of which can be modeled accurately. A new parallel pro-