Satellite-Aware Rate Adaptation (SARA)

Updated 8 July 2026

SARA is a middleware designed for live streaming over LEO satellites, mitigating handover-induced outages through proactive playback control and ABR input modulation.
It integrates seamlessly with existing ABR algorithms by adjusting playback speed and scaling network measurements to pre-buffer before disruptions and quickly recover afterward.
Evaluations show that SARA significantly reduces rebuffering events while maintaining low latency and minimal bitrate loss, proving effective across multiple ABR families.

Satellite-Aware Rate Adaptation (SARA) is a middleware for live streaming over Low Earth orbit Satellite Networks (LSNs) that is designed to mitigate the disruptive effect of frequent satellite handovers on application-layer video delivery. Introduced in "Robust Live Streaming over LEO Satellite Constellations: Measurement, Analysis, and Handover-Aware Adaptation," SARA sits between a video player’s adaptive bitrate (ABR) logic and the LSN network stack, and combines outage awareness, playback-speed control, and ABR-input modulation to reduce rebuffering while preserving low latency and bitrate quality. The motivating observation is that existing live streaming platforms, as well as current state-of-the-art learning-based ABR algorithms even when trained on LSN traces, do not handle abrupt handover-induced network variations effectively (Fang et al., 18 Aug 2025).

1. Network setting and design objective

The immediate context for SARA is live streaming over LEO satellite constellations such as SpaceX’s Starlink and Amazon’s Project Kuiper. In this setting, the dominant systems problem is not ordinary stochastic throughput fluctuation alone, but frequent, abrupt network outages caused by satellite handovers. The measurement study reports that these handovers lead to frequent video rebuffering events, and that terrestrial-optimized ABR algorithms degrade substantially under such conditions (Fang et al., 18 Aug 2025).

SARA is therefore defined by three concurrent objectives: minimizing rebuffering during handovers, maintaining low latency, and preserving high bitrate quality. Its design reflects a specific mismatch between terrestrial-centric streaming control and LSN dynamics. Conventional ABR logic reacts to observed buffer and throughput, whereas LSN handovers require a mechanism that can act before disruption arrives and can compensate after disruption passes. This is why SARA combines proactive and reactive controls rather than treating bitrate selection as an isolated problem.

A common misconception is that satellite-aware adaptation in this setting is simply a more conservative bitrate selector. SARA is broader than that. It modulates playback speed, alters the buffer and throughput values perceived by the ABR algorithm, and uses outage awareness to encourage pre-buffering before a predicted outage and latency recovery afterward. In that sense, its target is the end-to-end live streaming loop rather than bitrate selection alone.

2. Middleware placement and ABR integration

SARA is introduced as a versatile and lightweight middleware layer that wraps around an existing ABR algorithm. It is explicitly ABR agnostic and can operate with a wide class of ABR algorithms, including buffer-based, throughput-based, and hybrid or learning-based methods. The defining integration property is plug-and-play operation: SARA requires no change in the ABR decision logic, only input modulation (Fang et al., 18 Aug 2025).

Operationally, the control loop proceeds at each chunk interval. SARA obtains the current buffer, throughput, and outage-prediction information; an optimizer then sets three control variables: playback speed, buffer scalar, and throughput scalar. The modulated values are passed to the ABR algorithm, whose native logic selects the chunk bitrate. SARA then adjusts player playback speed and retrieves and plays the selected chunk.

This architecture is significant because it preserves the implementation investment embodied in existing ABR methods. Buffer-based ABRs receive a scaled view of buffer occupancy, throughput-based ABRs receive a scaled view of estimated throughput, and learning-based ABRs can be augmented without retraining their core policy logic. The paper’s evaluation includes RobustMPC, Pensieve, BOLA, BBA, DynamicDash, and Merina, illustrating that the middleware is intended as a general augmentation layer rather than a replacement controller.

3. Core control variables and optimization formulation

SARA exposes three principal control variables. The first is playback speed, denoted by $\beta_k \in [0.95, 1.03]$ at chunk $k$ . Before a predicted outage, SARA may slow playback, with $\beta_k < 1$ , to accumulate extra buffer. After an outage, it may speed playback, with $\beta_k > 1$ , to reduce end-to-end latency, described as live latency to broadcaster (LtB) (Fang et al., 18 Aug 2025).

The second and third variables are the ABR guidance scalars. SARA provides a buffer scalar $s^b_k$ and a throughput scalar $s^r_k$ , both functioning as multiplicative factors in $[0,1]$ . These shift the ABR algorithm’s perception of current operating conditions so that, ahead of an expected outage, bitrate selection becomes more conservative. The purpose is not to alter the ABR’s internal objective directly, but to steer its input state toward outage-resilient behavior.

The outage-aware buffer planning is formalized through a predicted-download calculation. The number of chunks downloadable before outage is given as

$\theta_{k} = \min \left\{ \left\lfloor \frac{\xi_k o_k^t}{b_{k}\alpha} \right\rfloor, \left\lfloor \frac{o_k^t}{\alpha} \right\rfloor \right\}$

where $\xi_k$ is measured throughput, $o_k^t$ is time until predicted outage, $k$ 0 is current bitrate, and $k$ 1 is chunk duration.

The post-outage buffer sufficiency constraint is presented as

$k$ 2

where $k$ 3 is outage duration, $k$ 4 is current buffer in seconds, and $k$ 5 is a safety margin for post-outage buffer, empirically set, with $k$ 6 seconds given as an example. When this constraint cannot be met, SARA minimizes the stall duration

$k$ 7

The quality-of-experience objective is expressed as

$k$ 8

where $k$ 9 is a video quality function, $\beta_k < 1$ 0 is rebuffer time, $\beta_k < 1$ 1 is extra live latency, and $\beta_k < 1$ 2 are weights. SARA’s optimizer solves

$\beta_k < 1$ 3

subject to the buffer, download, and playback constraints. Because the optimization variables are continuous, the implementation uses Particle Swarm Optimization (PSO), with each particle representing a candidate tuple $\beta_k < 1$ 4.

4. Outage prediction and handover awareness

SARA’s satellite awareness derives from explicit anticipation of handover-induced outages. The middleware can harness network or server insight and can also integrate with ML-based predictors to forecast imminent handovers and outages. The supporting example given in the paper is an Informer-based forward predictor that ingests Starlink network data, orbital element data (TLE), and weather information to forecast outage timing and duration (Fang et al., 18 Aug 2025).

This predictive component matters because LSN outages are abrupt relative to the control timescale of ordinary ABR adaptation. If outage timing is known or predicted, SARA can slow playback before the disruption to accumulate buffer, can bias bitrate decisions downward by modulating perceived buffer and throughput, and can then speed playback after the disruption to catch up toward the live edge. The design is thus explicitly handover-aware rather than merely throughput-aware.

In the reported experiments, the predictor achieved $\beta_k < 1$ 5 accuracy and recall of $\beta_k < 1$ 6. These figures indicate that prediction is useful but imperfect. A plausible implication is that SARA is designed to remain functional under prediction uncertainty because it combines multiple weakly coupled controls: playback-speed modulation, conservative ABR guidance, and buffer planning. Its robustness therefore does not depend on perfect outage forecasting.

5. Evaluation across ABR families

SARA was evaluated using measured Starlink outage traces embedded in a simulator configured for live-stream video, across six ABR algorithms: RobustMPC, Pensieve, BOLA, BBA, DynamicDash, and Merina. The central reported result is that SARA reduced average rebuffering time by $\beta_k < 1$ 7, slightly improved latency by $\beta_k < 1$ 8, and incurred only an overall bitrate loss of $\beta_k < 1$ 9 (Fang et al., 18 Aug 2025).

Metric	Reported outcome
Rebuffering time	Reduced by an average of $\beta_k > 1$ 0
Rebuffering events	Reduced by $\beta_k > 1$ 1 on average
Live latency (LtB)	Slight improvement of $\beta_k > 1$ 2
Bitrate	Overall loss of $\beta_k > 1$ 3
Non-1x playback fraction	Increased by $\beta_k > 1$ 4

The paper further reports that, for BBA specifically, rebuffering time reduction reached $\beta_k > 1$ 5. It also notes that the fraction of time spent at non- $\beta_k > 1$ 6 playback speed increased by $\beta_k > 1$ 7, but because the speed range is constrained to $\beta_k > 1$ 8, these changes are described as typically imperceptible.

Several qualitative trends are emphasized. First, the outage-prediction and pre-buffering mechanism can absorb handover-induced outages and keep playback smooth. Second, learning-based ABRs such as Pensieve still benefit significantly from SARA because, without SARA, they react too slowly to abrupt LSN outages. Third, the benefit is reported as consistent across major ABR archetypes. The BBA example illustrates the mechanism directly: without SARA, Starlink handovers induce frequent and severe stalling; with SARA, slower playback and aggressive buffering before a predicted outage can allow playback to ride through the handover without visible impact.

The acronym “SARA” is not unique in the literature. In "Exploiting Regional Differences: A Spatially Adaptive Random Access," Dong Min Kim and Seong-Lyun Kim use SARA to denote a spatially adaptive random access scheme in which each transmitter updates its transmit probability according to the ratio of received SINR to target SINR. That paper explicitly does not directly mention satellite-aware rate adaptation or satellite networks as a focus (Kim et al., 2014). This distinction is important because the live-streaming SARA is an application-layer middleware for ABR augmentation, whereas the earlier SARA is a distributed random-access mechanism.

Within satellite communications more broadly, SARA belongs to a larger family of adaptation techniques that condition transmission decisions on channel dynamics, delay structure, or predicted link opportunity. "Adaptive Transmission Techniques for Mobile Satellite Links" studies multi-layer coding in the forward link and open-loop adaptation in the return link, relying on physical-layer abstraction to increase average spectral efficiency while keeping outage probability under a threshold (Arnau et al., 2013). "Adaptive Network Coding Schemes for Satellite Communications" proposes physical layer aware adaptive network coding and coded modulation for time-variant channels with different RTTs, and highlights the trade-off between energy efficiency and delay-throughput gains under high erasure conditions (Gharsellaoui et al., 2017).

The term Satellite-Aware Rate Adaptation is also used in adjacent satellite downlink settings outside live video streaming. "Scalable Data Transmission Framework for Earth Observation Satellites with Channel Adaptation" describes a downlink framework for multi-spectral imagery that couples semantic prioritization of changed pixels with predicted link capacity, reducing the number of transmitted bits by up to $\beta_k > 1$ 9– $s^b_k$ 0 compared to raw or naive transmission while maintaining high semantic fidelity (Bui et al., 2024). "Learning to Compress and Transmit: Adaptive Rate Control for Semantic Communications over LEO Satellite-to-Ground Links" applies RL-based adaptive image transmission with SNR prediction and an on-board transmission-queue model, reporting nearly $s^b_k$ 1 qualified frames with zero packet loss under realistic overpass conditions (Luo et al., 11 May 2026).

Taken together, these papers show that “satellite-aware rate adaptation” is not a single algorithmic pattern but a class of adaptation strategies spanning PHY, MAC, network coding, semantic communication, and application-layer streaming. In that broader landscape, SARA in live streaming is distinguished by its handover-aware, ABR-agnostic middleware design and by its use of small playback-speed adjustments and input scalars to shape application behavior around predictable LEO handovers.