Letting Physicists be Physicists, and other Goals of Scalable Solver R&D

Abstract

Computational enabling technologies must offer both high-level abstractions in the language of their intended user community and detailed access to their powerful, composable innards for developers and prototypers. This can be achieved through a multilevel interface with robust default settings for a host of tuning parameters, the knobs for which can be exposed on demand. Drawing upon collaborations between DOE’s Integrated Software Infrastructure Center for scalable solvers (“TOPS”, www.tops-scidac.org) and projects in fusion energy, accelerator design, and QCD, we illustrate the necessity of a providing a rich set of linear solvers under a common high-level interface, in order to progress beyond mere feasibility to performance and portability. While many customers can be adequately served with well defined multilevel software interfaces, we emphasize from this same set of SciDAC collaborations the importance of cross-disciplinary human interaction to discover altogether better abstractions (e.g., one fully coupled nonlinear problem, rather than a sequence of operator-split, linearized problems). We also mention prospects for applying tools from machine learning to the automatic tuning of complex iterative solvers to particular tasks that arise often, e.g., in production runs of an application code.