Skip to main content

Low-Latency Operation

OpenIPC Station uses bounded queues. USB reads stay posted, WFB/RTP data is processed as soon as it is recoverable, and compressed frames go to WebCodecs without a playback jitter buffer.

Browser path​

WebUsbReceiverSession owns the nusb WebUSB endpoint and keeps four bulk-IN transfers in flight by default. Each completion is processed inside Rust/WASM and its buffer is recycled immediately. Only completed video frames, selected route payloads, and diagnostics cross into JavaScript.

This removes the raw USB transfer round trip through JavaScript. One final WASM-to-JavaScript copy remains for compressed data supplied to WebCodecs; decoded surfaces remain in the browser.

Desktop path​

The Tauri receiver keeps native USB and WFB processing on its worker thread. Video frames use a raw Tauri channel with a compact binary header and Annex-B payload. Counters and control information use the lower-rate JSON event. This avoids base64 expansion and JSON serialization for video frames.

Queue policy​

If the WebCodecs decode queue grows beyond the station limit, the decoder is reset and the receiver waits for the next IDR frame instead of displaying stale footage. RTP reordering is disabled by default and should only be enabled when the link actually delivers out-of-order packets.

Audio starts with a short scheduling cushion. Recording adds browser media encoding work and can increase latency on slower machines.

Source-side settings​

For wfb-rs, use zero tunnel aggregation when interactive latency matters. WFB primary fragments are emitted immediately; FEC parity is generated when a source block completes. Smaller FEC blocks reduce recovery wait at the cost of less parity efficiency.

The transmitter encoder remains part of the latency budget. Low-delay encoder settings, no lookahead, no B-frames, short GOPs, and frequent IDR frames are required for low glass-to-glass latency.

Measuring latency​

Diagnostics cover the local path:

USB completion -> Realtek parse -> WFB/FEC -> RTP frame
    -> IPC/WASM boundary -> decoder queue -> decoded output -> canvas

These measurements identify receiver bottlenecks but are not a true glass-to-glass number. That requires a transmitter capture timestamp or a synchronized on-screen timestamp.