“Is the remote desktop fast?” is rarely answered by one number. The user experiences the time from mouse movement to visible change, and that path spans network, capture, encode, transfer, decode and render. Splitting the stages is the only way to know where optimization belongs.
Direct first, relay as fallback
LANs, home networks and many corporate paths can hole-punch, so media never touches a relay. Only when symmetric NAT or a restricted network blocks a direct path does RemoteControl grant a relay, selected by region and zone. The benefit is direct: one less forwarding hop, one less queue, and less relay bandwidth.
The relay forwards ciphertext
RemoteControl relays do not inspect session content. Quotas, shaping and metering happen on the forwarding path while the payload stays end-to-end encrypted. For teams that need audit, the control plane records connection metadata, usage and policy outcomes, not screen content.
Explicit budgets
We use stage budgets instead of a single headline metric:
- LAN input-to-photon p95 ≤ 80ms; WAN p95 ≤ 150ms.
- Same-host relay p99 forwarding target ≤ 1ms at 1 Gbps with 64 KiB frames.
- Input is separated from the video queue, so clicks and keys keep flowing under congestion.
Those targets feed regression gates. A regression over threshold needs a fix or an explicit rationale.
Transport choices
Raw TCP and WebSocket remain supported transports. Where network and policy allow, QUIC, HTTP/3, WebTransport and HTTP/2 fallback are available. The point is not to enable everything, but to make each path observable, limitable and safe to fall back from.