HyperRouter: Lessons Learnt from Building an L4 Load...
summary
Describes building a Level 4 software load balancer that uses DPDK over eBPF/XDP for a user‑space kernel bypass data plane to hit about 10M pps per CPU core, implements direct server return to avoid forwarding large responses, and uses BGP path prepending (AS_PATH manipulation) for safer node degradation instead of withdrawals. Control plane is a decentralized peer-to-peer control plane with gossip and formal verification (which found a retry bug), plus local health checks, shuffle sharding to cut blast radius, and CUJ‑driven observability.
AI-written summary · about
chapters (13)
- 0:00HyperRouter intro
- 1:45Huawei Cloud scale
- 3:10Global load balancing role
- 4:57Level 4 load balancing
- 7:25HyperRouter high performance
- 9:53HyperRouter architecture
- 13:51Kernel bypass trade-offs
- 18:48Self-healing isolation
- 21:16Control plane decentralized
- 26:13Operational shuffle sharding
- 28:42Observability with critical user journeys
- 32:59Key lessons
- 33:19Q&A
related talks
37mEfficient and Flexible Virtual Machine Networking Through eBPF
47mInto the Heart of the eBPF Hive!
18mTGW: Operating an Efficient and Resilient Cloud Gateway at Scale
22mLayer 3 networking with BGP in hyperscale DCs
33mCloud Native Load Balancing From Scratch
35mGrout # a Graph Router Based on DPDK
32mLeveraging open source dynamic routing to implement a spine/leaf in datacenter
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs