基于CUDA的IB-LBM并行算法设计与性能分析

Design and Performance Analysis of CUDA-Based Parallel Algorithm for IB-LBM

摘要: 浸入边界–格子Boltzmann方法(IB-LBM)已成为模拟可变形颗粒–流体相互作用的重要数值工具。然而，在大规模复杂流固耦合问题的数值模拟中，普遍存在内存访问效率低与计算冗余严重等挑战，限制了数值仿真的计算性能与网格规模。为此，本文基于CUDA并行计算平台，对IB-LBM算法进行了并行算法优化设计，其中提出了双指针稀疏矩阵存储策略以解决全矩阵存储模式下无效节点更新导致的计算冗余问题。在此基础上，以多孔介质流与分叉管道流为例开展测试，并对比分析不同孔隙率下两种存储模式的计算效率与显存占用情况。结果表明，在孔隙率小于0.85时，所提出的优化策略可显著减少无效内存访问，平均计算性能提升20%以上，同时有效降低显存消耗。本文研究结果为大规模复杂流固耦合系统的高效数值模拟提供了可行的并行优化方案。

Abstract: The Immersed Boundary-Lattice Boltzmann Method (IB-LBM) has become a powerful tool for simulating deformable particle-fluid interactions. However, its application to large-scale fluid-structure interaction problems is often bottlenecked by inefficient memory access and significant computational redundancy. To address these issues, this paper presents a CUDA-based parallel algorithm optimization of the IB-LBM, and proposes a double-pointer sparse matrix storage strategy to eliminate the redundancy caused by invalid node updates inherent in conventional full matrix storage schemes. The presented algorithm is evaluated using porous media and bifurcation flow benchmarks, comparing the computational efficiency and GPU memory footprint of the sparse and full storage modes. The numerical results show that when porosities below 0.85, the proposed strategy significantly reduces invalid memory accesses, yielding a performance improvement of over 20% while substantially lowering memory usage. This work provides an effective parallel optimization scheme for high-performance simulations of complex, large-scale fluid-structure interaction problem.