Abstract
This article discusses problems at the software design and implementation level that transcend any particular subfield of computational mathematics, but which are important to all subfields. The discussion includes desirable attributes of numerical software that are independent of the underlying algorithms, ways in which software with these attributes exploit arithmetic characteristics of the host computer, advantages to the gathering of numerical software into thematic packages and the problems involved in such work, and challenges in testing numerical software. A nontechnical overview of the subject is followed by a technical discussion which includes suggestions for research activities within DOE.
Publisher
Association for Computing Machinery (ACM)