Accelerating GNN Training on CPU+Multi-FPGA Heterogeneous Platform
Lists
Only you see this
Abstract
Training Graph Neural Networks (GNNs) has become time consuming as the graphs grow larger. Thus, many works have been proposed to accelerate GNN training on multi-GPU platforms. Though GPUs feature high computation power, training GNNs on GPU suffers from low resource utilization. We propose to accelerate GNN training on a CPU+Multi-FPGA heterogeneous platform. By utilizing the customizable hardware resources on the FPGAs, we instantiate multiple hardware kernels with optimized data access pattern and memory organization. The optimized hardware kernels can efficiently access graph-structured data and thus achieve high training performance. However, training GNN with multiple FPGAs also leads to high FPGA-to-FPGA communication overhead and workload imbalance. We develop optimized graph partitioning techniques to minimize FPGA-to-FPGA data communication, and develop a task scheduler to balance the workload among the FPGAs. Compared with the state-of-the-art GNN training implementation on a multi-GPU platform, our work achieves up to 24.7\(\times \) bandwidth efficiency; this superior efficiency enables our work to achieve up to 3.88\(\times \) speedup and 7.18\(\times \) energy efficiency using much less compute power and memory bandwidth than GPUs.
BibTeX
@inproceedings{lin2022accelerating,
author = {Yi-Chien Lin and Bingyi Zhang and Viktor K. Prasanna},
title = {{Accelerating GNN Training on CPU+Multi-FPGA Heterogeneous Platform}},
booktitle = {CARLA},
year = {2022},
doi = {10.1007/978-3-031-23821-5\_2},
}Loading related content...