📝 SummarXiv

Abstract

A tridiagonal matrix algorithm (TDMA), Pipelined-TDMA, is developed for multi-GPU systems to resolve the scalability bottlenecks caused by the sequential structure of conventional divide-and-conquer TDMA. The proposed method pipelines multiple tridiagonal systems, overlapping communication with computation and executing GPU kernels concurrently to hide non-scalable stages behind scalable compute stages. To maximize performance, the batch size is optimized to strike a balance between GPU occupancy and pipeline efficiency: larger batches improve throughput for solving tridiagonal systems, while excessively large batches reduce pipeline utilization. Performance evaluations on up to 64 NVIDIA A100 GPUs using a one-dimensional (1D) slab-type domain decomposition confirm that, except for the terminal phase of the pipeline, the proposed method successfully hides most of the non-scalable execution time-specifically inter-GPU communication and low-occupancy computation. The solver achieves ideal weak scaling up to 64 GPUs with one billion grid cells per GPU and reaches 74.7 percent of ideal performance in strong scaling tests for a 4-billion-cell problem, relative to a 4-GPU baseline. The optimized TDMA is integrated into an ADI-based fractional-step method to remove the scalability bottleneck in the Poisson solver of the flow solver (Ha et al., 2021). In a 9-billion-cell simulation on 64 GPUs, the TDMA component in the Poisson solver is accelerated by 4.37x, contributing to a 1.31x overall speedup of the complete flow solver.