360,000 servers or 4,500? Route, don't mix

transcript

Fourteen people on a video call. Somebody has to combine all that video. Pick the classic way, and you need 80 times more servers.

Option one: every phone sends to every other. Arjun uploads 13 copies, from a cab. Option two, the old conference bridge. The server mixes. Option three: the server just forwards packets.

Mixing decodes every camera. Then encodes a separate picture for every receiver. And adds delay. Forwarding reads each packet's headers, and copies it to whoever wants it. It never decodes a pixel. About a millisecond added.

Peak hour: about 20 million people in meetings. 30 terabits a second out. At 10 gigabits per forwarding server, about 3,000 servers. Zoom says it keeps 50 percent spare. So, about 4,500. Mixing would decode 12 million cameras, and encode 20 million pictures. 360,000 servers.

But here's the twist. A forwarder never needs to see a pixel. So it can relay end-to-end encrypted media it can't read. A mixer never could.

Eighty times fewer servers. Less delay. Each receiver picks its own view. Route. Don't mix.

Full breakdown in the lecture.