NetChain: Scale-Free Sub-RTT Coordination
summary
NetChain moves the replicated coordination KV store into programmable switches (PISA) to cut client-visible latency to about 9.7 microseconds and to push throughput far beyond server-based systems — a prototype uses four Barefoot Tofino switches and four commodity servers, stores ~100k keys with up to 128‑byte values, and remains unsaturated by 82M packet/sec of client traffic. It implements chain replication in the data plane, uses per-key sequence numbers to handle reordering, and a two‑phase switch install for fast failure recovery.
AI-written summary · about
chapters (11)
- 0:00Coordination challenges
- 1:44In-network approach
- 3:25Chain replication design
- 6:55Nana-chain implementation
- 8:20Programmable switch basics
- 10:19Key-value storage on switches
- 12:21Packet routing protocol
- 14:11Handling network issues
- 15:48Switch failure recovery
- 17:49Evaluation and results
- 19:08Conclusion and Q&A
related talks
17mPushing the Limits of In-Network Caching for Key-Value Stores
15mThe nanoPU: A Nanosecond Network Stack for Datacenters
28mRock around the clock (synchronization
25mShoal: A Network Architecture for Disaggregated Racks
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
55mPractical TSN; using NetChan for distributed PoC
17mIn Link We Trust: BFT at the Speed of CFT using Switches
12mDebugging Transient Faults in Data Centers using Synchronized Network-wide Packet