[00934] Slab LU, a sparse direct solver for heterogeneous architectures
Session Time & Room : 1E (Aug.21, 17:40-19:20) @E704
Type : Contributed Talk
Abstract : This talk describes a scalable sparse direct solver for linear systems that arise from the discretization of elliptic PDEs in 2D or 3D. The scheme uses a decomposition of the domain into thin subdomains, or ``slabs''. The general framework is easier to optimize for modern heterogeneous architectures than
than traditional multi-frontal schemes. Crucial to the scalability, are novel randomized algorithms that recover structure from matrix-free samples and reduce the dimensionality of large dense matrices.