Mg-fim: a multi-gpu fast iterative method using adaptive domain decomposition

Applying the latest parallel computing technology has become a recent trend in Eikonal equation solvers. Many recent studies have focused on parallelization of Eikonal solvers for multithreaded CPUs or single GPU systems. However, multi-GPU Eikonal solvers are largely underresearched owing to their complexity in terms of data and task management. In this paper, we propose a novel adaptive domain decomposition method to realize an efficient implementation of the block-based fast iterative method on multiple GPUs. The proposed method progressively expands the computational domain assigned to each GPU to maximize load balancing and employs a locality-aware clustering algorithm to minimize inter-GPU communication overhead. We also propose various low-and high-level optimization techniques for GPU computing, such as overlapping CPU and GPU computation and inter-GPU data transfer using multiple CUDA streams. Thus, we effectively circumvent performance issues in the na{\i}\"ve parallelization using a regular decomposition method. The proposed method scales up to 6.6\times for eight GPUs. We demonstrate that our efficient parallel implementation of the proposed method achieves an improvement in runtime performance under various experimental setups.