Composable software for distributed GPU linear algebra

Pratik Nayak (Technical University of Munich)

Abstract: Modern simulation codes increasingly demand linear solvers that are simultaneously distributed, GPU-resident, and flexible enough to mix and match algorithmic components. In this talk, we argue that composability, the design choice that lets any linear operator play any role, whether as a solver, preconditioner, smoother, or coarse solver, is what makes this tractable in practice.
We will illustrate this through Ginkgo, an open-source linear algebra library built around a uniform LinOp abstraction. After motivating the composability thesis, we will walk through Ginkgo’s distributed functionality: on-device assembly of distributed matrices, GPU-aware sparse matrix–vector products with neighborhood communicators that reduce MPI overhead at scale, and a preconditioner suite, including block-Jacobi, ISAI, Schwarz, algebraic multigrid, and BDDC that all share the same composable interface. We will close with a brief look at applying these tools to large-scale cardiac electrophysiology simulations.