Alibaba HPN: A Data Center Network for Large Language Model Training
summary
A purpose-built 2-tier, dual-plane datacenter fabric connects segments of up to 1,000 accelerators per segment and scales to 130k accelerators using 51.2 Tbps switches, cutting hash collisions and narrowing path selection so periodic 400 Gbps elephant flows find stable routes. Hardware includes custom 51.2 Tbps switches, 200/400G optics and programmable NICs; software includes Echo CCL and SOLAR transport, yielding up to 96% linear scaling efficiency, 98.22% availability, and minute-scale failure recovery.
AI-written summary · about
chapters (16)
- 0:00Alibaba HPN 7.0 intro
- 1:06LLM training network challenges
- 3:56HPN 7.0 key solutions
- 4:40Two-tier topology benefits
- 5:21Multi-topology segment advantages
- 6:25Dual-plane topology redundancy
- 7:31Front-back separation performance
- 8:15Scaling to 130k accelerators
- 8:58Dual NIC non-stacked solution
- 9:40Hardware for stability
- 10:44Echo and SOLAR transport
- 11:46NS container network virtualization
- 12:06AI Pi job scheduling
- 12:49Network monitoring failure detection
- 13:31Production performance gains
- 14:35Future optimizations collaboration
related talks
27mPredictable Network
17mMeta’s Network Journey to Enable AI
05mKeynote: Innovation of the Network Infrastructure at Alibaba
23mNetworking for GenAI Training and Inference Clusters
18mKeynote: Enterprise Data Center Migration to Hyperscale – Zaid Ali Kahn
22mResearch SuperCluster(RSC): Accelerating AI Research at Scale
18mResiliency at Scale: Managing Google’s TPUv4 Machine Learning Supercomputer
21mDC Networks for Generative AI