Experimental options

Everything on this page changes a path that is off by default. Read its limits before turning it on.

Forward error correction

FEC adds redundancy so a receiver can reconstruct lost packets without asking for a retransmission. On a lossy wireless link this sounds exactly like what you want.

Do not enable FEC for browser clients

Chrome’s H.264 decoder needs contiguous RTP sequence numbers. The FEC scheme available here shares the sequence number space with media, which produces gaps the decoder reacts to by demanding keyframes, repeatedly. The result is a keyframe storm that is far worse than the packet loss it was meant to fix.

This is an upstream Chromium behaviour. The alternative scheme that would avoid it is not viable here. The conclusion from the investigation was: no FEC for browser clients, and revisit it for a native client.

The settings exist and are documented for completeness:

Setting	Default	Purpose
QUASAR_FEC_MODE	derived	off, fixed or auto.
QUASAR_FEC_PERCENTAGE	0	Redundancy percentage in fixed mode.

An auto mode exists that negotiates FEC at zero percent up front and ramps it only when it observes loss, then unwinds when the link is clean again. It is the least bad way to use FEC if you are going to. It is still not recommended for browsers.

FEC also competes with your video for bitrate, because the congestion controller sees the combined total.

Intra refresh

Instead of periodic full keyframes, intra refresh sends a rolling wave of intra-coded blocks. Recovery from packet loss is then bounded by one refresh cycle rather than by waiting for the next keyframe, and the periodic bitrate spike that a keyframe causes goes away.

QUASAR_INTRA_REFRESH=1
QUASAR_INTRA_REFRESH_PERIOD=60

A period of 60 is the validated setting. It matched continuous refresh for loss recovery at lower bandwidth cost. 0 means continuous.

Constraints:

• Vulkan encoder only. QUASAR_ENCODER=vulkan is required.
• A patched image is required. On stock GStreamer the property does not exist and the agent silently does nothing.
• H.264 only. HEVC and AV1 ignore it, with a warning.
• Pair it with more than one slice. The refresh cycle length follows the effective slice count.

Measured behaviour on an impaired link: zero visible hitches across the tested network shapes, where the same test with refresh off produced stalls under loss.

Resolution and frame-rate ABR rungs

QUASAR_ABR_LADDER_RESOLUTION and QUASAR_ABR_LADDER_FPS are shipped and actuated, but default off while they accumulate broader hardware validation. They require QUASAR_ABR_MODE=smooth and apply to the next session. See Adaptive bitrate for the codec and launch rate limits.

Things that look like knobs and are not

apps.default_vram_mb is deprecated. It was never enforceable, because there is no VRAM cgroup controller, and live telemetry replaced it.

The in-compositor NV12 mode was removed. It was an alternative to the current zero-copy path, it was not a win, and it blocked a compositor upgrade.

A native client does not exist in a usable form. Work is under way in a separate repository.

Before you change anything

Take a measurement first, change one thing, measure again. Every default here came from a measured comparison, and the differences involved are frequently smaller than the run-to-run variation you get from not controlling the test.

Diagnostics covers how to take a comparable measurement.