Abstract: Memory bandwidth is often the limiting factor for scientific simulations on modern exascale hardware, including DOE's latest GPU-based supercomputers. Matrix-free representations of high-order finite element-like operators are less expensive than sparse matrices, with respect to both the FLOPs and bytes of memory transfer required for a matrix-vector product. However, utilizing matrix-free implementations requires rethinking data structures and algorithms throughout the simulation stack. libCEED is an extensible library that provides a portable algebraic interface and optimized matrix-free implementations for a variety of modern supercomputer architectures. In the first part of the talk, we explore the design constraints and practical considerations for libCEED. Ratel is a solid mechanics library built using libCEED and PETSc to take advantage of the performance of matrix-free operators on modern supercomputer architectures while providing an API that legacy applications can leverage to migrate to modern hardware. In the second part of the talk, we tour some of the simulation capabilities available in Ratel, a solid mechanics simulation library developed using libCEED and PETSc. Matrix-free operators necessitate iterative solvers and good preconditioners along with supporting data structures, which offers a rich area of research, especially for applications that have typically relied upon sparse direct solvers. Ratel and libCEED are open source, with graduate students contributing key functionality to these libraries as part of their research. In the final part of this talk, we discuss how open source development best practices can facilitate research and mentoring for graduate students.

Bio: Jeremy L Thompson is a Research Software Engineer in the department of Computer Science at the University of Colorado Boulder. He develops the Ratel and libCEED libraries as part of CU Boulder's PSAAP center. After graduating from the US Air Force Academy, he served as an advanced weapon systems analyst, testing the accuracy and reliability of aircraft based nuclear weapons. He previously served on the faculty at the US Air Force Academy. He completed his PhD in applied mathematics from the University of Colorado Boulder in 2021. His research interests include preconditioning for high order finite element-like operators and sustainable open source scientific software development.