CRI: CRD-- An On-Demand Test Problem Server

University Of Texas At Austin, Austin TX

Investigators

Abstract

Algorithm researchers require realistic test problems to validate their work. However, these researchers may not have access to the applications that would produce such test problems. Consequently, they have to rely on test problems stored on web sites, which may be too limited in variety and unrealistically small, or on hand-coded test problems, which will not be realistic. This research concerns the construction of a test problem server, that is, a server that has a number of realistic research and production codes installed, with the facility to deliver to the algorithm researcher the test problems these codes produce. With such a server, the researcher can specify in advance the desired size and properties of the test problems, benefitting the quality of the algorithm research. Algorithms such as for linear systems, graph theory, or eigenvalue computations, are often developed in mathematics and computer science departments, without the researcher having access to the sort of applications that these algorithms will benefit from. Consequently, algorithm researchers have to rely on the limited inventory of test problems on popular archive sites such as Matrix Market, or they have to code their own test problems, which are unlikely to be realistic. This project constructs a server that will have a number of research and production codes installed, with the facility to serve up generated problems to the user. These codes can be run on demand, with specified problem inputs controlling the problem size and numerical properties. Thus, researchers have access to test problems that are both wide ranging in variety and finely graded in numerical properties. This is of benefit to the quality and robustness of the algorithms developed.