Yeti: Stateless and Generalized Multicast Forwarding
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Yeti implements generalized multicast by moving forwarding state into packets: a centralized controller generates controller-generated labels that encode multicast distribution graphs and an ingress router tags packets, while simple label-processing logic at core routers adds/removes labels and forwards or duplicates packets. Evaluated on a NetFPGA prototype (Verilog/P4) and simulations over 14 ISP topologies with one million sessions, Yeti eliminates core-router state, sustains high packet rates, and cuts label overhead ~65% on average vs the closest label scheme, requiring only one ingress update versus ~48 for OpenFlow.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:32Online Watching of Live Events
- 1:46Multicast Saves Bandwidth
- 2:20Routing Objectives: Traffic Engineering
- 2:35Routing Objectives: Service Chaining
- 3:25Problem Statement: Generalized Multicast
- 3:49Challenges of Generalized Multicast Forward
- 5:37Proposed Solution: Yeti
- 6:32Overview of Yeti
- 7:37Label Types in Yeti
- 8:24Creating Yeti Labels
- 9:48Yeti Keeps Needed Labels Only
- 10:20More Details in the Paper
- 10:56Evaluation in Simulation
- 11:54Yeti vs BIER-TE: Label Overhead
- 13:16Yeti vs OpenFlow vs LIPSIN: Scalability
- 14:50Evaluation in a Testbed
- 15:53Summary
related talks
22mTransparent MultiNIC Routing for Large AI Models
18mTGW: Operating an Efficient and Resilient Cloud Gateway at Scale
19mEnabling Programmable Transport Protocols in High Speed NICs
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
16mState-Compute Replication: Parallelizing High-Speed Stateful Packet Processing
39mManaging (..) networking flows using Erlang
26mFlowBlaze: Stateful Packet Processing in Hardware
15mCost-effective and Reliable Global Internet Peering with Programmable Switches