eXpressSFU: Toward Super-Scalable Video Conferencing with SmartNICs
summary
eXpressSFU re-architects an SFU by decoupling fast-control and data planes and offloading packet operations to a commodity BlueField-2 SmartNIC, implementing a fast control plane that enforces bitrate decisions and a NIC data plane that does replication, encryption/decryption, and I/O. Measured microbenchmarks show packet cloning 4×, decryption 8×, encryption 3× improvements; end-to-end packet processing latency drops 8×, enabling supports 3× more users (up to ~400 concurrent in large rooms versus software SFUs) while cutting CPU compute power by up to 60%.
AI-written summary · about
chapters (13)
- 0:00eXpressSFU introduction
- 2:511080p video cost
- 3:33Modern SFU architecture
- 5:43SFU latency challenges
- 7:51SmartNIC acceleration
- 8:35SmartNIC offloading challenges
- 9:39Control and data plane
- 11:46System implementation
- 12:07Evaluation setup
- 12:49Latency and scalability
- 13:54Power efficiency gains
- 14:16Key contributions
- 14:36Q&A
related talks
30mJitsi Videobridge: the state of the art SFU that powers Jitsi Meet
01h 00mCommSec D1 - An Isolated Data Center Security Policy Model Using SmartNICs
08mdevelops the Distributed UPF with NVIDIA Bluefield-3 DPU
19mSwCC: Software-Programmable and Per-Packet Congestion Control in RDMA Engine
19mDisaggregating Stateful Network Functions
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
15mUnified Representor with Large Scale Ports
15mThe nanoPU: A Nanosecond Network Stack for Datacenters