TopSim: an extensible plugin-based framework for solving general-purpose large-scale engineering problems in a parallel computational environment

Abstract

This paper presents TopSim, an extensible C++ framework for developing numerical simulations in parallel and distributed environments. The framework’s contribution is architectural rather than algorithmic, arising from the integration of three key architectural features within a unified execution model: (1) a service-oriented plugin architecture configured at runtime, enabling code reuse and the replacement of numerical methods without modifications to the framework kernel; (2) an integrated data structure, TopS, which simultaneously provides a compact representation of the topological model and serves as a shared communication protocol among plugins; and (3) the encapsulation of distributed-memory programming tasks, such as mesh communication, attribute synchronization, and distributed linear algebra operations, through TopSim’s distributed-memory backbone, namely ParTopS, TopMat, and TopVec. The framework’s consistency, extensibility, and distributed performance are evaluated through mechanics-oriented benchmarks, including linear and geometrically nonlinear structural analyses, a topology optimization example based on a three-dimensional hexagonal-prism (honeycomb) finite-element discretization, and distributed scaling experiments. In the distributed algebraic microbenchmarks, TopSim’s native distributed linear algebra (DLA) algorithms achieved lower absolute elapsed times than the corresponding PETSc DLA algorithms across most of the configurations tested. In the end-to-end finite-element benchmarks, absolute elapsed time depended on the per-core workload: the TopSim-native configuration performed better when the workload exceeded 10K elements per core. In terms of scalability, the TopSim-native configuration exhibited weaker strong scaling but consistently better weak scaling than the TopSim-PETSc configuration. Overall, the results demonstrate that TopSim can incorporate new simulation capabilities with minimal modifications to the plugin architecture while maintaining competitive distributed performance.