Adaptive Code Collage: A Framework to Transparently Modify Scientific Codes

Abstract

Legacy scientific codes are often re-purposed to fit adaptive needs, such as to dynamically alter parameters to improve convergence behavior, or to switch algorithms at runtime for greater accuracy of modeling. Given a legacy scientific code, how can we make it adaptive without making changes to the original source program(s)? We present an approach-Adaptive Code Collage (ACC)-to achieve this goal using function call interception in a language-neutral way at link time. ACC transparently 'catches' function calls and redirects them so that an existing program can be made adaptive without causing a significant performance overhead. We demonstrate the appliction of ACC to designing adaptive SOR algorithms for solving linear systems and to improving the performance, stability, and accuracy of GenIDLEST, a large parallel computational fluid dynamics code.

Function call interception techniques

“…The adaptive code collage (ACC) framework is a compositional tool for implementing program behavior adaptations on top of nonadaptive existing code . It realizes a static way of FCI by instrumenting the existing code with a hook to the framework APIs at the assembly level, where different traditional imperative languages can be translated into the same language semantics and can be integrated together thereby achieving programming language independence.…”