This paper describes work in progress to develop a standard for interoperability ariiong high-petforniance scientific coniponents. This research sterns front growing recognition that the scientific coniniunity needs to better manage the coniplexity of ntultidisciplinuiy simulations and better address scalable petforniance issues on parallel and distributed architectures. Driving forces are the need for fast connections among components that perform numerically intensive work and for parallel collective interactions among cornponetits that use multiple processes or threads. This paper focuses on the areas we believe are niost crucial in this context, naiizely, an intetface definition language that supports scientipc abstractions for specifying coriiponent interfaces and a ports connection model for specifiing component interactions.

The 0-1 multiple knapsack problem appears in many domains from financial portfolio management to cargo ship stowing. Methods for solving it range from approximate algorithms, such as greedy algorithms, to exact algorithms, such as branch and bound. Approximate algorithms have no bounds on how poorly they perform and exact algorithms can suffer from exponential time and space complexities with large data sets. This paper introduces a market model based on agent decomposition and market auctions for approximating the 0-1 multiple knapsack problem, and an algorithm that implements the model (M(x)). M(x) traverses the solution space rather than getting caught in a local maximum, overcoming an inherent problem of many greedy algorithms. The use of agents ensures that infeasible solutions are not considered while traversing the solution space and that traversal of the solution space is not just random, but is also directed. M(x) is compared to a bound and bound algorithm (BB) and a simple greedy algorithm with a random shuffle (G(x)). The results suggest that M(x) is a good algorithm for approximating the 0-1 Multiple Knapsack problem. M(x) almost always found solutions that were close to optimal in a fraction of the time it took BB to run and with much less memory on large test data sets. M(x) usually performed better than G(x) on hard problems with correlated data.

Abstract.With the increasing complexity and interdisciplinary nature of scientific applications, code reuse is becoming increasingly important in scientific computing. One method for facilitating code reuse is the use of components technologies [16,17, 91, which have been used widely in industry. However, components have only recently worked their way into scientific computing [2, 1, 11, 181. Language interoperability is an important underlying technology for these component architectures. In this paper, we present an approach to language interoperability for a high-performance parallel, component architecture being developed by the Common Component Architecture (CCA) group.' Our approach is based on Interface Definition Language (IDL) techniques [6]. We have developed a Scientific Interface Definition Language (SIDL), as well as bindings to C and Fortran. We have also developed a SIDL compiler and run-time library support for reference counting, reflection, object management, and exception handling (Babel). Resuits from using Babel to call a standard numerical solver library (written in C) from C and Fortran show that the cost of using Babel is minimal, where as the savings in development time and the benefits of object-oriented development support for C and Fortran far outweigh the costs.

Executive SummaryWe are developing new software component technology for high-performance parallel scientific computing to address issues of complexity, re-use, and interoperability for laboratory software. Component technology enables cross-project code re-use, reduces software development costs, and provides additional simulation capabilities for massively parallel laboratory application codes. The success of our approach will be measured by its impact on DOE mathematical and scientific software efforts. Thus, we are collaborating closely with library developers and application scientists in the Common Component Architecture forum, the Equation Solver Interface forum, and other DOE mathematical software groups to gather requirements, write and adopt a variety of design specifications, and develop demonstration projects to validate our approach.Numerical simulation is essential to the science mission at the laboratory. However, it is becoming increasingly difficult to manage the complexity of modern simulation software. Computational scientists develop complex, three-dimensional, massively parallel, full-physics simulations that require the integration of diverse software packages written by outside development teams. Currently, the integration of a new software package, such as a new linear solver library, can require several months of effort.Current industry component technologies such as CORBA, JavaBeans, and COM have all been used successfully in the business domain to reduce software development costs and increate software quality. However, these existing industry component infrastructures will not scale to support massively parallel applications in science and engineering. In particular, they do not address issues related to high-performance parallel computing on ASCI-class machines, such as fast in-process connections between components, language interoperability for scientific languages such as Fortran, parallel data redistribution between components, and massively parallel components. While industrial component systems do not directly address scientific computing issues, we leverage existing industry technologies and design concepts whenever possible.Since the mid-year start of this project in FY99, we have focused on the needs of seamless language interoperability in a high-performance environment. Computational scientists are routinely hindered in code re-use by differences in programming languages; for example, a solver library written in C++ can be difficult to call from an applications code written in C or Fortran. Our approach adopts the industry practice of using an Interface Definition Language (IDL) to describe component interfaces in a language-independent manner. We have developed an IDL for scientific applications (SIDL) that focuses on the unique needs of the scientific domain as compared to the business world. We have also created tools that use SIDL descriptions of software components to generate automatically language bindings and code that allows the component to be called easily from ...