LGNNIC: SmartNIC-Based Architecture to Accelerate Large-Scale GNN Training
A new study on arXiv (2608.07733) reveals LGNNIC, a cutting-edge system design that employs SmartNICs to cut down on communication delays when training distributed Graph Neural Networks (GNNs). As graphs grow larger, relying on a single CPU-GPU setup isn't practical anymore, leading to a push for distributed solutions across remote memory nodes. However, training can become bogged down by network traffic between nodes. LGNNIC addresses this by offloading key preprocessing tasks to SmartNICs near the memory nodes, which helps lower the data sent to compute nodes and eases congestion. The study also introduces two techniques using SmartNICs to improve performance, tackling scalability challenges in GNNs that are relevant in various fields like social network analysis and chip design.
Key facts
- LGNNIC is a novel inter-node system architecture for distributed GNN training.
- It leverages SmartNICs co-located with remote memory nodes.
- SmartNICs offload key preprocessing tasks to reduce data transfer volume.
- The approach alleviates inter-node network congestion during training.
- Two complementary techniques are executed on SmartNICs during training.
- GNNs are used in natural sciences, social network analysis, chip design, and recommendation systems.
- The paper is available on arXiv with identifier 2608.07733.
- The architecture targets modern systems with existing SmartNIC configurations.
Entities
Institutions
- arXiv