Adaptive bitrate

Adaptive bitrate measures the connection continuously and moves the encoder’s target bitrate to match. It is on by default in smooth mode, and for most deployments that is the end of the story.

Why it exists

A fixed bitrate is wrong almost all the time. Set it high and any congestion turns into dropped packets and visible corruption. Set it low and you waste a good connection.

ABR watches round trip time, packet loss and the congestion controller’s bandwidth estimate, and moves the target as conditions change.

The two modes

Set with QUASAR_ABR_MODE, globally in deploy/.env or per host under Admin, Hosts, Settings.

smooth, the default

Tuned for how a stream feels rather than for maximising bitrate.

The key behaviour is on the way down. When the encoder is already saturated, a non-emergency reduction is capped at 12.5 percent per step with a seven second wait between steps, and the congestion controller alone cannot force a large cliff. The theory is that a large sudden drop is more visible than a slightly too-high bitrate held for a few more seconds.

The emergency path is preserved. A confirmed network congestion window still triggers a fast single-step descent.

Measured against protective under induced congestion: presentation interval variation at the 95th percentile fell from about 69 ms to about 19 ms, and freezes fell from 14 to 2. On a clean path the two modes are indistinguishable.

protective

The older governor. One step down when it sees trouble, gradual ramp back up. More willing to drop hard and fast.

Worth trying if smooth behaves oddly on your hardware, particularly if the encoder speed ladder causes problems.

off

Fixed bitrate at whatever the tier specifies. The congestion controller stays attached for telemetry but never moves the encoder.

Reasonable on a dedicated wired LAN where the bandwidth genuinely is not going to change. Not reasonable over WiFi.

Turning it off

QUASAR_ABR_MODE=off

Or the older aliases, which still work:

QUASAR_ABR=0

The floor

ABR will not reduce below a floor, so a bad moment cannot leave you at an unwatchable bitrate.

By default the floor is 30 percent of the tier’s bitrate, clamped to at least 500 kbps. Set it explicitly with QUASAR_ABR_FLOOR_KBPS, or change the ratio with QUASAR_ABR_FLOOR_RATIO.

The encoder speed ladder

In smooth mode, sustained encoder saturation escalates the encoder’s speed setting rather than only cutting bitrate. Trading some compression efficiency for encode time is often the better deal when the GPU is the bottleneck rather than the network. It unwinds after four healthy windows.

Disable it with QUASAR_ABR_LADDER=0.

Resolution and frame-rate rungs are opt-in

QUASAR_ABR_LADDER_RESOLUTION and QUASAR_ABR_LADDER_FPS are implemented but default off. Both require smooth mode and apply to the next session. The FPS rung halves 120 fps to 60 and restores it; a 60 fps launch has no lower FPS rung. The resolution rung moves through the session’s supported external-size ladder. Support depends on the active codec and image; Vulkan HEVC cannot resize live.

Hardware only, really

ABR is validated on hardware encoders. On software encoding it barely functions, because openh264 cannot saturate a 1080p stream at 8 Mbps, so the congestion controller never observes congestion to react to.

If you are on software encoding, ABR is not the reason your stream is poor.

NVENC can overshoot

On high-entropy content NVENC sometimes produces more bits than the target ABR set. The controller compensates, but a stream that persistently sits above its setpoint on busy scenes is a known behaviour rather than a misconfiguration.

Watching it work

Admin, Sessions, session detail plots the ABR setpoint against the congestion controller’s estimate. The gap between the two is the smoothing that smooth mode is adding.

A setpoint sawtoothing rapidly means the controller is fighting something. Look at loss and round trip time in the same window before assuming it is the controller’s fault.

Fine tuning

There are roughly fifteen further settings covering the governor’s hysteresis: smoothing factor, deadband, maximum step sizes, dwell times, cliff guard fractions. Every default matches the value that used to be hardcoded, so leaving them alone reproduces the tested behaviour exactly.

Do not touch them without a measurement to compare against. See Diagnostics for how to take one. The full list is in Environment variables.