Abstract:

This chapter provides an introduction to the Portable Extensible Toolkit for Scientific Computing (PETSc), a scalable library designed for solving partial differential equations (PDEs) and other large-scale numerical problems. It outlines PETSc's hierarchical, object-oriented class structure, which is built upon abstract interfaces to provide flexibility and run-time polymorphism. The text details core components, including distributed data structures like vectors (Vec) and matrices (Mat), and the Data Management (DM) class for organizing data on various grids. It then explores the suite of solvers: KSP for linear systems, SNES for nonlinear systems, and TS for time-dependent problems (time-steppers). The chapter emphasizes PETSc's strong support for MPI-based parallelism and concludes with a practical example of solving a 2D Poisson equation, comparing the performance trade-offs between a direct (MUMPS) and an iterative (GMRES) linear solver.

An Introduction to the PETSc Framework for Scientific Computing

• Introduction: What is PETSc? PETSc is a comprehensive toolkit for the numerical solution of PDEs. It is designed with a hierarchical class structure, starting from low-level data structures (vectors, matrices) and building up to high-level solvers for complex problems. Its use of abstract interfaces allows algorithms and data storage formats to be modified at runtime without recompiling.
• 34.1: Getting Started: To use PETSc, you must install it and set the $PETSC_DIR and $PETSC_ARCH environment variables. PETSc is built on MPI and provides convenient wrapper functions, PetscInitialize and PetscFinalize, for handling both PETSc and MPI setup and teardown. It also includes built-in error handling via the PetscCall function.
• 34.2: Core Data Structures: PETSc provides Vec classes for distributed 1D vectors and Mat classes for distributed sparse or dense matrices. Data is populated using VecSetValues and MatSetValues and then distributed across MPI ranks using assembly calls (VecAssemblyBegin/End). For complex geometries, the DM (Data Management) object helps manage data distribution on structured (DMDA) and unstructured (DMPlex) grids.
• 34.3.1: Linear Solvers (KSP): The KSP object provides a unified interface for solving linear systems of the form Ax = b, supporting both direct and iterative methods. For iterative solvers, performance often depends on a preconditioner, which is managed by a PC object.
• 34.3.2: Nonlinear Solvers (SNES): The SNES class is used to solve nonlinear systems of equations (F(x) = 0). It supports numerous algorithms, with Newton's method as the default. These solvers internally use a KSP object to solve the linear system required at each nonlinear iteration.
• 34.3.3: Time-Steppers (TS): The TS class provides algorithms for time-dependent problems, such as systems of Ordinary Differential Equations (ODEs) and Differential-Algebraic Equations (DAEs). It includes a range of methods from fully explicit to fully implicit, many with adaptive time-stepping.
• 34.4: Parallelism Model: PETSc's primary parallelism model is distributed memory using MPI, where it hides most of the low-level message-passing details from the user. It also has support for OpenMP and GPUs (CUDA, HIP, Kokkos).
• 34.5: Key Takeaway: Direct vs. Iterative Solvers: A practical example solving a 2D Poisson equation compares a direct solver (MUMPS) with an iterative solver (GMRES).
  • Direct Solvers: Deterministic and robust but often suffer from poor scalability and high memory requirements, making them less suitable for extreme-scale problems.
  • Iterative Solvers: Generally more efficient in time and memory for large problems but may require a good preconditioner to guarantee convergence.
  • For the example's small problem size, the direct solver was ~4x faster but used nearly 2x the memory of the iterative solver. Error: value error Invalid operation: The response.text quick accessor requires the response to contain a valid Part, but none were returned. The candidate's finish_reason is 1.