Unified Representor with Large Scale Ports
summary
Disables per-VF/SF software representors and instead routes miss packets into a single uplink proxy representor, with the hardware writing the origin port into packet metadata as a source port ID cookie. Implementation removes per-port RX/TX queue setup and polling, cutting hugepage use from ~12 GB (1k representors) / ~800+ MB in small tests to about 332 MB, and raises forwarding from ~20 Mpps to ~27.5 Mpps when adjusting xburst. Works with standard DPDK switchdev and NIC offload rules.
AI-written summary · about
chapters (7)
- 0:00Introduction and background
- 3:20Challenge with large scale ports
- 6:14Optimization approach
- 9:06Rule table changes
- 11:17Memory and performance gains
- 13:25Test results and summary
- 14:55Q&A
related talks
31mFlow-based Packet Processing Framework on DPDK and VPP
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
51mAccelerate containers and VMs networking with vDPA
25mEvolution of Direct Server Return (DSR) implementation for containerized applications
33mSUPERp: Scale UP EtherRnet Protocol
22mTransparent MultiNIC Routing for Large AI Models
19mDisaggregating Stateful Network Functions