Generating Customized Sparse Eigenvalue Solutions with Lighthouse
Ramya Nair, Sa-Lin Bernstein, Elizabeth R. Jessup, Boyana Norris · International Multi-Conference on Computing in Global Information Technology · 2014
Sparse eigenvalue problems arise in many areas of scientific computing. A variety of high-performance numerical software packages including many different eigensolvers are available to solve such problems. The two main challenges are finding the routines that can correctly solve the problem and implementing the desired solution accurately and efficiently using the appropriate software package. In this paper, we describe an approach that addresses these issues by intelligently identifying the sparse eigensolver that is likely to perform the best for given input characteristics and by generating a code template that uses that solver. The results are delivered to users through Lighthouse, a novel interface and search platform for users seeking highperformance solutions to linear algebra problems. This paper describes the development of the approach with a focus on the analysis of sparse eigensolvers in SLEPc and their integration into Lighthouse. Keywords–expert systems; sparse eigensolvers; machine learning.