Legion: Automatically Pushing the Envelope of Multi-GPU System for Billion-Scale...
summary
Claims an end-to-end multi-GPU system that trains billion‑node GNNs on a single machine by combining hierarchical graph partitioning to minimize cache replication, a hotness‑aware unified cache for topology+features, and an automatic cache‑size optimizer that maximizes throughput. Built on DGX/TGX platforms and evaluated on real billion‑scale datasets, Legion cuts PCIe traffic dramatically, prevents GPU OOMs that break other caches, raises cache hit rates, and reports up to ~4.3× speedup versus prior cache systems.
AI-written summary · about
chapters (5)
related talks
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
31mThe Evolution of Multi-GPU Inference in vLLM
26mEfficiently serving LLMs at scale.
38mExploring Distributed Caching for Faster GPU Training with NVMe, GDS, and RDMA
24mHow to Eliminate the I-O Bottleneck & Continuously Feed the GPU While Training in the Cloud
44mArchitecture Effects When Interacting with GPU Technologies
13mFrom Silicon to Scale: Training at Scale
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...