Optimizing Thermal Lattice Boltzmann Method on an MT-3000 Processor with Neo-Heterogeneous Strategies

Abstract

The Lattice Boltzmann Method (LBM) is a computational fluid dynamics method for simulating fluid flows with the benefits of locality and simplicity, making it ideal for parallel computing and complex flow simulations. This study focuses on developing a specialized Double Distribution Function (DDF) LBM software framework optimized for the MT-3000, a novel heterogeneous processor, to facilitate thermal incompressible flow simulations. To improve LBM’s performance on the complex multi-zone architecture of MT-3000, this paper introduces several innovative strategies. Firstly, a temporal fusion optimization strategy is implemented. This strategy involves postponing the temperature field calculations during time steps, efficiently decreasing the time overhead. Furthermore, we present “Pencil-H”, a novel pipelined algorithm meticulously designed to harness the unique capabilities of the MT-3000, thereby enhancing computational efficiency and communication effectiveness. Additionally, an architecture-aware multi-level parallelization algorithm is proposed, tailored to maximize the computational capabilities of the MT-3000. The effectiveness of these optimization strategies has been thoroughly validated through extensive bench-marking tests. These validations have shown remarkable performance enhancements, including a significant acceleration factor of 32.02X when compared to using 16 CPU cores. Notably, the optimized code demonstrated high-fidelity simulation capabilities for thermal incompressible flows, achieving 61.61% of the theoretical maximum performance defined by the roofline model.