지리적 확장을 위한 SFU 캐스케이딩

A single SFU (Selective Forwarding Unit) forwards each participant's media to others. But one SFU has limits: CPU, network egress, and the latency of being far from some users.

Cascading SFUs link multiple SFUs together to break through that ceiling.

Instead of every participant connecting to one central SFU, participants connect to a nearby SFU. Those SFUs then forward streams to each other.

• Local users connect to a local node.
• Nodes exchange only the streams that remote users actually need.

A user in Tokyo connecting to an SFU in Virginia suffers high round-trip latency. With cascading, the Tokyo user hits a Tokyo SFU, which relays over a fast backbone to other regions.

This cuts first-hop latency, the part users feel most.

Without cascading, a remote region pulls every stream individually across the ocean. With cascading, a stream crosses the expensive inter-region link once, then fans out locally.

SFUs connect to each other much like clients do, but optimized for trunk traffic. They negotiate which simulcast layers to relay so a large 4K stream is not blindly copied everywhere.

A key optimization: an SFU only requests a remote stream when a local participant subscribes to it. If nobody in Tokyo is watching a Berlin speaker, that stream never crosses the link.

• Full mesh: every SFU peers with every other. Simple, but N-squared links.
• Hub and spoke: regional SFUs feed a central relay. Fewer links, one extra hop.
• Tree: hierarchical, good for very large broadcasts.

Use geo-DNS or a latency probe so a joining client lands on the closest SFU. The signaling server records which SFU owns each participant for cross-node relay decisions.

If a regional SFU dies, its participants must migrate to a backup node. Health checks plus a participant registry let the system re-home users and re-establish relays quickly.

Combine cascading with simulcast: each publisher sends multiple quality layers. The cascade relays only the layer a remote viewer can use, saving trunk bandwidth on poor links.

• Sessions with hundreds of participants.
• Globally distributed audiences.
• Large webinars and town halls.

For small regional calls, a single SFU is simpler and cheaper.

Cascading SFUs chain regional nodes so users connect locally and streams cross inter-region links only when needed. This scales large global sessions while cutting latency and trunk bandwidth.