Bifrost: Alibaba's Next-Generation VPC Network with High-Performance Multipath...
summary
Bifrost moves reliable transport out of the VM into the VPC by inserting a new header and running a transport engine beside AVS on the SmartNIC, using RTT-aware multipath packet spraying to avoid hotspots, in-place guest reordering by reordering DMA descriptors so the guest sees ordered packets, and delayed bitmap ACKs with 4 ms RTO. It supports O(100k) concurrent connections per SmartNIC with 1 GB memory, achieves <1 s failure recovery, and reports up to 307× tail-latency reduction for Redis and 66× for Nginx in production deployment.
AI-written summary · about
chapters (11)
- 0:00Alibaba Cloud VPC intro
- 2:26VPC performance evolution
- 3:34Reliable transport motivation
- 6:00Network reliability challenges
- 7:25Bifrost architecture overview
- 9:09RTT-aware multipath spraying
- 11:20In-order delivery design
- 12:46Delayed ACK performance gains
- 14:13Memory efficiency in Bifrost
- 15:01Real-world deployment results
- 15:25Conclusion and future
related talks
19mBifrost: Analysis and Optimization of Network I/O Tax in Confidential Virtual...
18mTGW: Operating an Efficient and Resilient Cloud Gateway at Scale
18mThe Evolution of High Performance Networks: A Storage Perspective and Practices
19mBluebird: High-performance SDN for Bare-metal Cloud Services
05mKeynote: Innovation of the Network Infrastructure at Alibaba
27mPredictable Network
22mTransparent MultiNIC Routing for Large AI Models
27mAzure Accelerated Networking: SmartNICs in the Public Cloud
