UC Irvine

Physics-based simulation, Computational Fluid Dynamics (CFD) in particular, has substantially reshaped the engineering design process by dramatically reducing the experimental cost.The effectiveness of CFD is primarily driven by the power of High Performance Computing (HPC) systems. However, the CFD community's top-priority task has been to develop high-order accurate schemes for complex physical phenomena like turbulence, instead of designing algorithms towards better usage of the hardware. Moreover, the hardware is evolving at a much faster pace than the numerical methods. As a result, we currently lack efficient algorithms for the practical CFD schemes to fully leverage the HPC systems. In this dissertation, we propose high-performance algorithm designs to improve the performance and productivity of CFD solvers using multi-block structured grids on modern supercomputers.

First, to parallelize simulations, we need to partition and distribute the grid across processors.Current grid partitioners only factor communication metrics and ignore the underlying network and architectures. To achieve portable performance, we introduce a novel cost function combining the algorithmic metrics with the network properties. Based on the cost function, we propose new partitioning algorithms that minimize the communication cost while balancing the computation workload.

Second, the computation in CFD solvers is typically memory-bound and optimized with cache tiling. The state-of-the-art tiling techniques are mainly designed for single-block grids on shared memory systems, whereas multi-block grids distributed across many nodes are the norm for realistic CFD applications. In this thesis, we demonstrate how to tile distributed computation for multi-block grids. Furthermore, we propose novel pipelined algorithms to hide the communication cost which improves both the performance and strong scaling of CFD solvers.

At last but not least, constructing a Deep Learning(DL) surrogate for CFD can further improve its productivity in engineering design.The surrogates must be able to predict for solutions across different boundary conditions and geometries unseen during training. In this dissertation, we design transferable DL surrogates for Boundary Value Problems (BVP) resembling the CFD simulations for steady-state flow. Furthermore, we propose an iterative inference method that transfers the prediction across arbitrary domain sizes and shapes.