Predictable Network
summary
Alibaba Predictable Network uses a Host-Network Fusion design to deliver predictable performance for large AI clusters, running large-scale RDMA for five years and claiming GPU clusters scale linearly with cluster size. Infrastructure uses SONiC-based single-chip white boxes (single ASIC supporting 256×100G, moving to a 51.2T chip), P4-based edge boxes, an EPF router system that cuts network emergencies ~10×, HPCC congestion control, a SOLAR UDP-based protocol, and x-link SDKs on tens of millions of endpoints. SRv6 migration uses a hybrid TE overlay.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 3:00Evolution of network technology over two decades
- 5:52From product to software networks
- 7:42Predictable networks next
- 9:53White box for DC and backbone
- 11:58Host-network co-design
- 14:29RDMA for predictable performance
- 19:08SRv6 migration challenges
- 23:26Backbone migration to SRv6
- 25:16Open source contributions
- 26:44Summary
related talks
15mAlibaba HPN: A Data Center Network for Large Language Model Training
05mKeynote: Innovation of the Network Infrastructure at Alibaba
17mMeta’s Network Journey to Enable AI
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
18mThe Evolution of High Performance Networks: A Storage Perspective and Practices
21mDC Networks for Generative AI
08mKeynote: The Challenge and Opportunity to Build a Transparent Network
19mScaling AI Network with DSF