The International Conference for High Performance Computing, Networking, Storage and Analysis
Generating Customized Eigenvalue Solutions Using Lighthouse.
Authors: Luke Groeninger (University of Colorado Boulder), Ramya Nair (University of Colorado Boulder), Sa-Lin Cheng Bernstein (University of Chicago), Javed Hossain (University of Colorado Boulder), Elizabeth R. Jessup (University of Colorado Boulder), Boyana Norris (University of Oregon)
Abstract: Eigenproblems are a fundamental part of science and engineering. Indeed, at least one thing you have done today was made possible in part by solving an eigenproblem. Whether it was using a search engine or speech recognition software, driving across a bridge, or flying on an aircraft, eigenproblems are now responsible for many everyday services on which you depend. Scientists and engineers across many disciplines must thus rely on high-performance numerical libraries to solve eigenproblems quickly and efficiently. However, exploiting the latest high-performance computational techniques for solving eigenproblems remains difficult. Our work addresses this problem by enabling users to find and use customized routines for solving eigenproblems without the steep learning curve traditionally associated with HPC libraries. We present new contributions in three areas: taxonomies of dense and sparse eigensolvers available in LAPACK and SLEPc; user interfaces for searching the taxonomies; and performance-based recommendation of solvers based on machine-learning analysis.