30 Tbps
of live audio and video at peak
About 11M people sit in meetings at peak. Assume each sends about 1 Mbps and receives 1.5. 16.5 Tbps leaves the servers, so egress dominates.
Ten cameras. Who carries them?
Bad: Mesh: send to everyone
Each uploads 9 × 1.5 = 13.5 Mbps. Calls collapse around four people.
Good: SFU: upload once
It forwards packets without decoding, each sized to its tile.
Fibre, Wi-Fi and a train. Who gets which video?
Asha uploads 720p, 360p and 180p, about 2.15 Mbps. The SFU forwards each receiver the largest layer its tile and bandwidth allow.
Wi-Fi and 4G drop packets. Re-asking a sender across an ocean takes too long.
The trick
The receiver NACKs the SFU, which keeps about a second of every stream. The repair crosses one short hop.
Five in India, four in Toronto. Where does the call live?
With one SFU in Mumbai, 36 streams cross the ocean. Cascade an SFU in each region and each stream crosses once, 9 in all.
200
Mbps of video per 100-person meeting, not 1,485
Every camera to everyone, even at 180p, is 100 × 99 × 0.15 Mbps. Forward each person one 25-tile page instead, and only the three loudest voices.
Behind a strict firewall? Walk down the ladder
- UDP straight to the SFU first
- UDP on port 3478 or 443
- TCP to a TURN relay on 443
- TLS on 443: looks like web traffic
Remember
Control over WebSockets, media over UDP through nearby SFUs. Upload simulcast once; forward each receiver what it can take; cascade SFUs across regions.