Energy-conscious bitrate selection and budget allocation for live video transcoding systems using deep reinforcement learning

초록

The growing demand for scheduled live broadcasts, such as those on platforms like Twitch, has transformed video delivery, requiring efficient transcoding strategies to meet real-time quality demands. These broadcasts rely on dynamic adaptive streaming over HTTP (DASH) to adapt video quality seamlessly to fluctuating network conditions. However, the computational demands of transcoding multiple bitrate versions pose significant challenges for energy efficiency, particularly in carbon-intensive environments where meeting strict energy budgets is essential. We propose a two-fold solution combining a deep reinforcement learning (DRL)-based framework and a greedy algorithm to optimize bitrate selection and energy budget allocation for live broadcasting. The DRL framework leverages an observation space that includes cumulative metrics for video quality and energy consumption, enabling it to dynamically adapt to uncertainties in popularity fluctuations and transcoding workloads. Using expected video quality as a reward signal, it determines the optimal bitrate versions to transcode. To complement this, we present a dynamic energy budget allocation algorithm that adjusts hourly energy budgets based on fluctuating access requests using a greedy algorithm, ensuring efficient energy use and workload management. Experimental results demonstrate that this combined approach enhances video quality by an average of 11.2 % (ranging from 2.6% to 20%) across varying energy budgets, achieving a sustainable balance between energy efficiency and service quality.

키워드