A comparative study of DASH representation sets using real user characteristics

Abstract

Adaptive streaming strategies over HTTP allow to serve heterogeneous video users with varying demands. By providing different encoded versions (representations) of each video sequence on the server, clients have the freedom to select a representation that best fits their needs. While the topic of selecting a representation based on a pre-defined set is covered very well in the literature, the problem of how to properly select the representation set stored at the main server is usually an overlooked challenge. In this work, we provide an analysis on how the choice of representations on the server impacts the clients' quality. This is achieved by conducting NS-3 based simulations with a total of 10k users and up to 300 concurrent DASH clients for several recommended sets (e.g., Netflix, YouTube, and Apple), and measuring the experienced quality over a timespan of 24 hours. The results show that under heavy load (at peak hours) there is still room for improvement.

Delay-Power-Rate-Distortion Optimization of Video Representations for Dynamic Adaptive Streaming

“…On the other hand, Ref. [10] validates the optimized representations obtained by solving the ILP in [4] in a practical scenario, by generating a 24hour streaming scenario based on YouTube traces and device statistics for Hulu and Netflix. Since the ILP formulation proposed in [4] and [10] can be viewed as a special case of problem P2 without the maximum encoding delay and power constraints, its performance is upper bounded by the baseline scheme 3).…”

“…Different works have been proposed recently to optimize the multiple representation selection for dynamic adaptive streaming [2], [4]- [10]. Most of these research efforts focus on the client-side adaptation algorithms in order to guarantee the quality of experience (QoE) of users for given encoded representations at the server, such as the live streaming rate adaptation method to support a smooth presentation while maintaining a small buffer size [5], the application layer probe-and-adapt rate adaptation approach driven by an estimate of the network dynamics [6], and the online rate adaptation algorithm in order to minimize the re-buffering phases [7].…”

“…In [4], the optimal representation set selection problem of adaptive streaming under the encoding rate constraint of the DASH server is proposed as an integer linear program (ILP), revealing the best coding parameter in terms of the bitrate and resolution for each representation. In [10], the optimized representations obtained by solving this ILP are further investigated and validated in a practical scenario, by generating a 24-hour streaming scenario based on YouTube traces and device statistics for Hulu and Netflix. These two works are rate-distortion efficient, capable of achieving the best overall video quality with the minimum cost of total encoding bitrate.…”

“…We implement the proposed framework on a 48-processor server with 252 GB of RAM using Linux 3.1 kernel, where each processor is an Intel Xeon CPU E5-2680 at a clock frequency of 2.50 GHz. We suppose that there are N = 10 users and their download bandwidth Bn is randomly distributed in the rate range of [1,10] Mbps. Three test video sequences (F = 3, Crowd Run, Tractor, and Sunflower) with 1080p resolution (1920 × 1080) [31] are selected as the source video streams to be encoded at the DASH server.…”

“…In addition, there are two remarks. First, the ILP formulations in [2], [4] and [10] do not include the rate control consideration. To make a "fair" comparison, we assume that the ground set of all possible representations is already pre-encoded with known bitrates, qualities and power consumptions when solving these ILPs.…”