Enabling End-to-End Accelerated Multiphysics Simulations in the Exascale Era Using PETSc

The Portable Extensible Toolkit for Scientific Computation (PETSc) library provides scalable solvers for nonlinear time-dependent differential and algebraic equations and for numerical optimization; it is used in dozens of scientific fields, and has been an important building block for many computational geoscience applications. Starting from the terascale era in the 1990s and continuing into the present dawn of the exascale era, a major goal of PETSc development has been achieving the scalability required to fully utilize leadership-class supercomputers. We will describe some of the algorithmic developments made during the era in which achieving inter-node scalability was the primary challenge to enabling extreme-scale computation, and then survey the challenges posed by the current era in which harnessing the abundant fine-scale parallelism within compute nodes -- primarily in the form of GPU-based accelerators -- has assumed at least equal importance. We will discuss how the PETSc design for performance portability addresses these challenges while stressing flexibility and extensibility by separating the programming model used by application code from that used by the library. Additionally, we will discuss recent developments in PETSc's communication module, PetscSF, that enable flexibility and scalable performance across large GPU-based systems while overcoming some of the difficulties posed by working directly with the Message Passing Interface (MPI) on such systems. A particular goal of this talk will be to go beyond simply describing the work performed to prepare PETSc and simulation codes that rely on it to run on exascale-class systems, but to enumerate the challenges we encountered and to share the essential lessons learned that can help other developers to prepare and optimize their high-performance scientific computing codes for the exascale era.