Reducing the Communication Overhead of Graph Neural Networks with SmartNICs Hardware

In distributed GNN (Graph Neural Network) training/inference, communication overhead has become a major bottleneck due to its costly message passing and aggregation operations. Different computation units need to exchange and synchronize their data which introduces latency and additional computation time. To reduce the communication overhead and improve the overall performance of GNN training/inference, we propose to design a SmartNICs hardware accelerated GNN framework. Specifically, we plan to study the following sub-problems: 1) how to partition large graphs on to different nodes to achieve load-balancing; 2) how to utilise the compute power of SmartNICs hardware for preprocessing and reducing the size of transferred data; 3) how to utilise the DRAM of SmartNICs to achieve better data locality to reduce communication overhead. We will work closely with Huawei researchers in this project.