Numerical Methods for PDEs: In Occasion of Raytcho Lazarov's 70th Birthday
- Stanimire Tomov, University of Tennessee
- Numerical Linear Algebra for Emerging Architectures: Challenges and Approaches
Numerical PDEs and their application in a wide range of science and engineering applications depend on the availability of high-performance linear algebra libraries; these applications will not perform well unless linear algebra libraries perform well. In the past, high-performance numerical library developers have taken for granted that each successive generation of microprocessors would make their old software run substantially faster. The emergence of the multicore and heterogeneous processor designs have brought this "free ride" to an end. Existing numerical libraries built on the old model are unable to effectively exploit these new architectures without radical modification. In this talk I will describe some of the current challenges and approaches in addressing this critical and highly disruptive situation in the area of dense linear algebra. Examples will be given with new algorithms, extending the LAPACK and ScaLAPACK libraries to efficiently support heterogeneous systems of multicore CPUs, GPUs, and/or coprocessors.