An on-demand Test Problem Server

Matrix algorithms, especially sparse matrix methods, play an important role in computational science. The researchers that develop such algorithms (such as linear system solvers, eigensolvers, mathematical optimization, and graph algorithsm) require test matrices that are, ideally, drawn from a large, varied, and realistic test collection. Unfortunately, existing collections (e.g., the Harwell-Boeing collection, Matrix Market, and the University of Florida collection) are often old, contain matrices that are unrealistically small for current computers, or are otherwise limited. We describe the design of a dynamic test collection based on solicited research and production codes, hosted at TACC (the Texas Advanced Computing Center of The University of Texas at Austin). These codes will be made available in the form of a test problem server, run on demand with properties (both size and problem parameters) specified by the user. Also, they will be used to generate a static test collection of large, precomputed test problems that can be searched through an extensive repertoire of properties. In both cases, the matrices will be made available via the web for download, as a TeraGrid Data Collection accessible at high speed from other TeraGrid resources, and directly from within codes running on TACC’s TeraGrid resources. Stored matrices will be accessible through a database that allows searching on a wide range of numerical features. Eijkhout et al. An on-demand Test Problem Server