Showing papers presented at "Computational Science and Engineering in 1998"

OpenMP: an industry standard API for shared-memory programming

Abstract: At its most elemental level, OpenMP is a set of compiler directives and callable runtime library routines that extend Fortran (and separately, C and C++ to express shared memory parallelism. It leaves the base language unspecified, and vendors can implement OpenMP in any Fortran compiler. Naturally, to support pointers and allocatables, Fortran 90 and Fortran 95 require the OpenMP implementation to include additional semantics over Fortran 77. OpenMP leverages many of the X3H5 concepts while extending them to support coarse grain parallelism. The standard also includes a callable runtime library with accompanying environment variables.

The NEOS Server

Abstract: The Network-Enabled Optimization System (NEOS) is an Internet based optimization service. The NEOS Server introduces a novel approach for solving optimization problems. Users of the NEOS Server submit a problem and their choice of optimization solver over the Internet. The NEOS Server computes all information (for example, derivatives and sparsity patterns) required by the solver, links the optimization problem with the solver, and returns a solution.

Accelerating fast multipole methods for the Helmholtz equation at low frequencies

Abstract: The authors describe a diagonal form for translating far-field expansions to use in low frequency fast multipole methods. Their approach combines evanescent and propagating plane waves to reduce the computational cost of FMM implementation. More specifically, we present the analytic foundations for a new version of the fast multipole method for the scalar Helmholtz equation in the low frequency regime. The computational cost of existing FMM implementations, is dominated by the expense of translating far field partial wave expansions to local ones, requiring 189p/sup 4/ or 189p/sup 3/ operations per box, where harmonics up to order p/sup 2/ have been retained. By developing a new expansion in plane waves, we can diagonalize these translation operators. The new low frequency FMM (LF-FMM) requires 40p/sup 2/+6p/sup 2/ operations per box.

IES/sup 3/: efficient electrostatic and electromagnetic simulation

Abstract: The fast integral equation solver presented in the paper is ideal for three-dimensional problems with smooth, far-field kernels. IES/sup 3/ solves electrostatic and electromagnetic problems in small circuit structures.

Applying NetSolve's network-enabled server

Abstract: The scientific community has long used the Internet for communication of e-mail, software, and papers. Until recently, there has been little use of the network for actual computations. This solution is changing rapidly and will have an enormous impact on the future. The NetSolve system described here has a significant role to play in these developments. The NetSolve project lets users access computational resources, both hardware and software, distributed across the network. Thanks to a variety of interfaces, users can easily perform scientific computing tasks without having any computing resources installed on their computers. Research issues involved in the NetSolve system include fault tolerance, load balancing, user interface design, computational servers, virtual libraries, and network based computing. As the project matures, several promising extensions and applications of NetSolve will emerge. We describe the project and examine some of the extensions being developed for NetSolve: an interface to the Condor system, an interface to the ScaLapack parallel library, a bridge with the Ninf system, and an integration of NetSolve and ImageVision.

The case for high-level parallel programming in ZPL

Abstract: Message passing programs are efficient, but fall short on convenience and portability. ZPL is a high level language that offers competitive performance and portability, as well as programming conveniences lacking in low level approaches. ZPL runs on a variety of parallel and sequential computers. We describe the problems with message passing and describe how ZPL simplifies the task of programming for parallel computers-without sacrificing efficiency.

A prescription for the multilevel Helmholtz FMM

Abstract: The authors describe a multilevel Helmholtz FMM (fast multipole method) as a way to compute the field caused by a collection of source points at an arbitrary set of field points. Their description focuses on the algorithm's mathematical basics, so that it can be applied to a variety of applications.

The accuracy of fast multipole methods for Maxwell's equations

Abstract: The multilevel fast multipole method can provide fast, accurate solutions to electromagnetic scattering problems, provided its users select the FMM degree and FMM cube size appropriately. The article discusses errors associated with truncating multipole expansions and methods for selecting an appropriate set of parameters.

The Fast Illinois Solver Code: requirements and scaling properties

Abstract: To solve large scale computing and scattering problems and to expand knowledge about iterative solvers, the authors developed the Fast Illinois Solver Code, which uses the multilevel fast multipole algorithm (MLFMA). They discuss FISC's memory requirements and CPU time and give some empirically derived formulas and charts. The authors also plot examples they used to obtain these conclusions.

Developing component architectures for distributed scientific problem solving

Abstract: Component programming models offer rapid construction for complex distributed applications, without recompiling and relinking code. This survey of the theory and design of component based software illustrates their use and utility with a prototype system for manipulating and solving large, sparse systems of equations.

Comparing job-management systems: the user's perspective

Abstract: The goal of job-management systems is to efficiently manage the distributed computing power of workstations, servers, and supercomputers. Effectively controlling these resources according to site-defined policies requires robustness, configurability, and allocation and scheduling mechanisms to maximize job throughput. There are now many highly functional job-management software packages in both the commercial marketplace and public domain. This comparison of job-management systems has highlighted the degrees to which these products fulfil end-user requirements. While all are highly functional, the user must decide whether ease of use is a more critical requirement than the extended functionality and flexibility found in the more complex systems. The relative costs of acquiring, installing, configuring, maintaining, and supporting these products are also important selection criteria that have merely been implied in this study and would be interesting considerations in future comparisons.

Numerical Computation: Methods, Software, And Analysis

Abstract: his book is a translated, revised, T and updated version of Computer-Numerik, published in German by Springer-Verlag in 1995. It's divided into two volumes because of the work's size (over 900 pages) rather than because of a natural division in content. Volume 1 contains material on model-ing, the basics of numerical methods and computers, numerical software, and interpolation. Volume 2 covers approximation theory, Fourier transforms , quadrature, numerical linear algebra (for both dense and sparse problems), nonlinear equations (including optimization), and random numbers. The book does not cover the solution of differential equations. In the spirit of earlier works by John R. Rice,l,* Ueberhuber emphasizes software and algorithmic aspects and does not go into as much mathematical detail as typical numerical analysis t e x t b o o b the book includes theorems, but they are usually stated without proof. It is therefore well-suited to mathematical-software and scientific-computing courses. Unlike Rice's books, this one contains no exercises, however. The book's strengths are its wide coverage, including many topics rarely treated in numerical analysis textbooks; its many pointers to software in major program libraries; its thorough bibliography ; and the many illustrative examples and figures. Material that is not easily found elsewhere includes computer memory systems (cache, virtual memory, and so on), performance measurement and optimization (including various loop transformations) for computer programs, how to access numerical software over the Internet (although this material already looks slightly dated, referring to Mosaic rather than Netscape and mentioning Archie rather than Web search engines), and thorough coverage of numerical integration, including the multi-dimensional case (in fact, Ueber-huber covers this topic in disproportionate detail, compared with most others). A few minor criticisms: Because of the wide coverage, the level of detail is patchy. For example, Ueberhuber includes the Jordan canonical form and principal vectors, but only defines, not explains, the practically important WY representation of products of House-holder transformations. The numerical example applying Lapack's condition estimator to Pascal-like matrices confuses a condition estimator's performance with the effects of rounding error (which are beyond the estimator's control). The index contains boldface page-number entries without defining their meaning, and contains no page ranges, even when an entry's coverage spreads across several pages. I found occasional typos (the definition of round to even, for example, is actually of round to odd) and a few statements that could potentially mislead or puzzle the reader (Horner's rule for …

Atomistic simulation of nanostructured materials

Abstract: Materials and devices with microstructures on the nanometer scale are revolutionizing technology, but until recently simulation at this scale has been problematic. The paper considers how developments in parallel computing are now allowing atomistic simulation using multiresolution algorithms, such as fast multipole methods. With these algorithms, researchers may soon be able to simulate applications up to one billion atoms.

Creating simulation capabilities

Abstract: On September 23, 1992, scientists and engineers of the US nuclear weapons complex in Nevada made the Earth move for the last time. The US government decided to stop developing new nuclear weapons now that the Cold War had ended. The need for underground nuclear tests, which had traditionally served primarily for testing and developing new designs, naturally decreased dramatically once the need for those new designs disappeared. But that last underground blast also marked the beginning of a new era in computer simulation. Besides testing new designs, underground testing had also answered critical questions about the safety, reliability, and performance of existing weapons. If anything, those needs would increase, not decrease, as the US nuclear weapons stockpile aged. So, as part of an effort to address those questions by developing a more thorough scientific understanding of the issues involved, the US Department of Energy propounded the Accelerated Strategic Computing Initiative (ASCI). ASCI's objective is to support high confidence assessments and stockpile certifications through high fidelity simulations. The ASCI program will create these simulation capabilities as part of the DOE's larger Stockpile Stewardship and Management program. The formidable nuclear weapons science issues involved in making this shift will require a significant investment in terms of scientific understanding and funding. The article discusses the steps the ASCI program will take in meeting that challenge.

Parallel languages for discrete-event simulation models

Abstract: The growing complexity of simulation models has promoted the increased use of parallelism to speed turnaround of large, detailed models of heterogeneous system. The article presents an overview of parallel simulation languages and briefly describes the Parallel Simulation Environment for Complex systems (Parsec), developed at UCLA.

Multilevel Evaluation of Integral Transforms on Adaptive Grids

Abstract: Fast numerical evaluation of integral transforms on an adaptive grid, i.e. using local grid refinement, requires an algorithm that relies on smoothness properties only of the continuum kernel, independent of its discrete form. The basic outline of such an algorithm was given in [6], where it was shown that already on a uniform grid this algorithm was more efficient than earlier fast evaluation algorithms [4, 5]. In this paper we outline its detailed formulation for the actual case of local grid refinements. Numerical results are presented for a model problem with a singularity. First it is shown that on a uniform grid this singularity dictates a much deteriorated relation between work and accuracy in comparison with the regular case (where accuracy is measured in terms of approximating the continuum transform, of course). Next we demonstrate that with the presented fast evaluation algorithm on a non-uniform grid one can restore the regular work to accuracy relation, i.e., obtain the same efficiency as for the case without a singularity.

Virtual rocketry: rocket science meets computer science

Abstract: The goal of the Center for Simulation of Advanced Rockets is comprehensive simulation of solid propellant rockets. Achieving this goal will require advancements in technical issues related to several scientific disciplines, in system integration, and in the computational infrastructure for supporting such large scale simulations. It requires a multidisciplinary team of engineers, physical scientists, and computer scientists to develop and implement the necessary mathematical models, algorithms, and software to build a virtual rocket. The authors outline their approach to this merger of rocket and computer sciences.

Concurrent computation and data visualization for spherical virus structure determination

Abstract: The authors present concurrent algorithms and programs for structure determination using structural information obtained through X-ray crystallography and electron microscopy. They introduce two interactive software systems, Emma and Tonitza, that support processing large data sets produced in structural biology experiments.

Models and high-performance algorithms for global BRDF retrieval

Abstract: The authors describe three models for retrieving information related to the scattering of light on the Earth's surface. Using these models, they've developed algorithms for the IBM SP2 that efficiently retrieve this information.

Parallel computing strategies for determining viral capsid structure by cryoelectron microscopy

Abstract: To calculate a full 3D structural model of a virus capsid, researchers analyzed cryoelectron micrographs that contain many randomly oriented images of the views. The authors use parallel computing techniques to improve the performance of the computational algorithms that determine each particle's orientation and generate the 3D model. This enhanced computational performance allows analysis of many more particles and a more precise determination of their orientations, letting researchers study important details of virus capsids at higher resolutions.

Programming languages for CSE: the state of the art

Abstract: To meet the diverse demands of building CSE applications, developers can choose from a multitude of programming languages. This survey offers an overview of available programming languages and the contexts for their use.

Fast Multipole Methods

Abstract: he past 10 to 15 years have witnessed widespread activity on a new class of fast numerical methods for applying linear operators. Collectively, they have come to be known as fast multipole methods (FMM). “Fast” is used in the sense of the fast Fourier transform (FFT): the cost of computing the result of the linear operation grows more slowly with problem size than it would if a naive representation were used. The new class differs from the old (consisting of FFT variants) in that the sparse decompositions used to render applications of the operators fast are fundamentally approximate. This is unlike the factorization of the FFT, which is exact. A more formal way to put this is that the new class of decompositions is based on analytic, rather than algebraic, properties of the operators. (The distinction is not, however, a practical limitation; correctly implemented, the attendant approximation errors can be made as small as necessary at a reasonable cost.) The new methods have begun to multiply the hardware power of computational science and engineering. Their widespread utility stems from the fact that the application of linear operators is often the dominant computational freight of physical simulation and modeling. This special issue aims to present assorted perspectives on the state of the art and research directions in this new field.

Fortran 90 in CSE: a case study

Abstract: An electromagnetics application demonstrates how Fortran 90's features facilitate the development of computational and science engineering (CSE) applications, and how a good compiler can generate efficient code from Fortran 90 source code, both in serial and parallel forms. The article examines the applicability of Fortran 90's two main new features, array language and abstract data types, to CSE. We collectively call scientific and engineering applications technical applications. Technical applications are typically numerically intensive computations, often manipulating large collections of data in the form of arrays. Technical programmers have long cherished the ability of Fortran implementations to produce efficient code for numerical computations.

Computational verifiability and feasibility of the ASCI program

Abstract: The Advanced Strategic Computing Initiative aims to replace physical nuclear weapons testing with computer simulations. This focuses much needed attention on an issue common to all Grand Challenge computation programs: how do you get confidence, as opposed to mere guidance or suggestion, out of a computer simulation? Making the simulation of some physical phenomena rigorous and immune to the usual discretization and rounding errors is computationally expensive and practised by a vanishing, small fraction of the HPCC community. Some of ASCIs proposed computing power should be placed not into brute force increases in particle counts, mesh densities, or finer time steps, but into methods that increase confidence in the answers it produces. This will place the debate over ASCI's validity on scientific instead of political grounds.

Language support for multidisciplinary applications

Abstract: The authors survey multidisciplinary application (MDA) characteristics and derive a corresponding set of language requirements, focusing on modular system design and multilevel parallelism. After evaluating the MDA computability of several languages and paradigms, they present Opus, the language they designed to support MDA requirements.

Using a fast multipole method to accelerate spline evaluations

Abstract: In considering the problem of interpolating scattered data using spline methods, the authors present a general framework for using the multipole method to accelerate spline evaluations. The article also illustrates the efficiency and accuracy of the fast multipole algorithm for the 2D vector spline.

Interleave in peace, or interleave in pieces

Abstract: Bit interleaving is a technique that sometimes can collapse a problem of high dimensional data to one of lower dimensional data. Our favorite use is going from two or three dimensions to one dimension. Of course, you lose something in this process because 2D is not the same as 1D. There is no "reasonable" one-to-one mapping between the two spaces. But sometimes, what you lose isn't all that important, and what you gain might make it worth doing. We show how we have used bit interleaving for two applications: finding the Hausdorff distance between sets (approximately) and calculating capacity dimension.

Teramac: pointing the way to real-world nanotechnology

Abstract: Researchers say that the ability to grow molecular sized electronic components-already being produced in research labs around the world-is expected to reach commercial economic and technical feasibility around the same time CMOS technology reaches its physical or economic limit (around 2010). A major goal of nanoscale technology is to build systems that incorporate a huge number of computational devices (researchers speak in terms of a mole, approximately 6/spl times/10/sup 23/). How will tomorrow's computer architects achieve the organization that allows the entire mass of devices to operate efficiently? Chemically assembled machines will certainly have to reproduce the arbitrary complexity that general purpose computation demands. Teramac, an experimental configurable computer built at the HP Lab, has demonstrated a viable solution to this problem. The powerful computer was originally designed to test different parallel computing architectures in the mid '90s, but has proven that a massively defect tolerant computer-that accommodates the uncertainty of "grown" computing devices-can be built.

Supercomputing: the next generation

Abstract: Just before November's Supercomputing'98 in Orlando, Florida, both Silicon Graphics and IBM claimed to have produced the world's fastest supercomputers. At the same time, Compaq announced a world record for sorting one terabyte of data on its latest supercomputer, and just four months earlier, Japan's NEC announced it had the world's fastest supercomputer. The hype of marketing and the fact that competing claims rarely use figures from the same benchmark test cause confusion. The paper discusses the technology trends behind the flurry of supercomputer development.

PathForward gets down to work [computer project]

Abstract: If the US Department of Energy's Accelerated Strategic Computing Initiative (ASCI) is successful in achieving its next teraflops milestone using commodity technology, the PathForward project will have made it possible. The ambitious ASCI goal is to achieve a threefold increase in computing performance every 18 months over a 10 year period using off-the-shelf parts. As one of the problem solving components of the ASCI program, PathForward's purpose is to leverage the technology of products currently being produced by computer companies. The author discusses the main aims of the PathForward project.