CBA: Contextual Quality Adaptation for Adaptive Bitrate Video Streaming (Extended Version) Bastian Alt+,∗, Trevor Ballard+,z, Ralf Steinmetzz, Heinz Koeppl∗, Amr Rizkz ∗Bioinspired Communication Systems Lab (BCS), {bastian.alt | heinz.koeppl}@bcs.tu-darmstadt.de zMultimedia Communications Lab (KOM), [email protected], {amr.rizk | rst}@kom.tu-darmstadt.de, Technische Universität Darmstadt, Germany Abstract—Recent advances in quality adaptation algorithms The problem of video quality adaptation is aggravated in leave adaptive bitrate (ABR) streaming architectures at a cross- Future Internet architectures such as Named Data Networking roads: When determining the sustainable video quality one may (NDN). In NDN, content is requested by name rather than either rely on the information gathered at the client vantage point or on server and network assistance. The fundamental problem location, and each node within the network will either return here is to determine how valuable either information is for the the requested content or forward the request. Routers are adaptation decision. This problem becomes particularly hard in equipped with caches to hold frequently-requested content, future Internet settings such as Named Data Networking (NDN) thereby reducing the round-trip-time (RTT) of the request where the notion of a network connection does not exist. while simultaneously saving other network links from redun- In this paper, we provide a fresh view on ABR quality adap- tation for QoE maximization, which we formalize as a decision dant content requests. Several attempts to make DASH-style problem under uncertainty, and for which we contribute a sparse streaming possible over NDN exist, e.g., [2], for which the Bayesian contextual bandit algorithm denoted CBA. This allows key difficulty is that traditional algorithms rarely play to the taking high-dimensional streaming context information, including strengths of NDN where the notion of a connection does not client-measured variables and network assistance, to find online exist. Throughput, for example, is not a trivial signal in NDN the most valuable information for the quality adaptation. Since sparse Bayesian estimation is computationally expensive, we as data may not be coming from the same source. develop a fast new inference scheme to support online video In this paper, we closely look at the problem of using adaptation. We perform an extensive evaluation of our adaptation context information available to the client for video quality algorithm in the particularly challenging setting of NDN, where adaptation. Note that our problem description is agnostic to we use an emulation testbed to demonstrate the efficacy of CBA the underlying networking paradigm, making it a good fit compared to state-of-the-art algorithms. to traditional IP-based video streaming as well as NDN. In essence, we consider the fundamental problem of sequential I. INTRODUCTION decision-making under uncertainty where the client uses net- Video streaming services such as Netflix, YouTube, and work context information received with every fetched video Twitch, which constitute an overwhelming share of current segment. In Fig. 1 we show a sketch where the client adap- Internet traffic, use adaptive bitrate streaming algorithms that tation algorithm decides on the quality of the next segment try to find the most suitable video quality representation based on a high-dimensional network context. We model the given the client’s networking conditions. Current architectures client’s decision on a video segment quality as a contextual use Dynamic Adaptive Streaming over HTTP (DASH) in multi-armed bandit problem aiming to optimize an objective conjunction with client-driven algorithms to adjust the quality QoE metric that comprises (i) the average video quality bitrate, arXiv:1901.05712v1 [cs.MM] 17 Jan 2019 bitrate of each video segment based on various signals, such (ii) the quality degradation, and (iii) the video stalling. as measured throughput, buffer filling, and derivatives thereof. One major challenge with incorporating high-dimensional In contrast, new architectures such as SAND [1] introduce network context information in video quality adaptation is network-assisted streaming via DASH-enabled network ele- extracting the information that is most relevant to the sought ments that provide the client with guidance, such as accurate QoE metric. We note that the interactions within this context throughput measurements and source recommendations. Given space become complicated given the NDN architecture, where the various adaptation algorithms that exist in addition to the network topology and cache states influence the streaming client-side and network-assisted information, a fundamental session. Our approach introduces a sparse Bayesian contextual question arises on the importance of this context information bandit algorithm that is fast enough to run online during for the Quality of Experience (QoE) of the video stream. video playback. The rationale behind the sparsity is that the given information, including network-assisted and client-side This work has been funded by the German Research Foundation (DFG) as measured signals such as buffer filling and throughput, consti- part of the projects B4 and C3 within the Collaborative Research Center (CRC) 1053 – MAKI. tutes a high-dimensional context which is difficult to model in + The first two authors equally contributed major parts of this article. detail. Our intuition is that, depending on the client’s network High-dimensional especially in NDN, and related work on contextual bandit Network packet-level context algorithms with high-dimensional covariates. assistance Significant amounts of research have been given to finding Client-side information Data streaming architectures capable of satisfying high bitrate and minimal rebuffering requirements at scale. CDN brokers such Sparsity enforcing as Conviva [3] allow content producers to easily use multiple Requests CDNs, and are becoming crucial to meet user demand [4]. Most important context features Furthermore, the use of network assistance in CDNs has Cross traffic received significant attention recently as a method of directly Quality Adaptation providing network details to DASH players. SAND [1] is an ISO standard which permits DASH enabled in-network ? entities to communicate with clients and offer them QoS in- Bitrate formation. SDNDASH [5] is another such architecture aiming Segment number to maintain QoE stability across clients, as clients without Fig. 1: A standard client-based and/or network-assisted ABR network assistance information are prone to misjudge current streaming model (black) with the proposed Context-based network conditions, causing QoE to oscillate. Beyond HTTP, Adaptation—CBA (dotted). In CBA, high-dimensional context the capabilities of promising new network paradigms such features from the network, along with client-side information, as NDN pose challenges to video streaming. The authors of undergo sparsity enforcement to shrink the impact of unim- [2] compare three state-of-the-art DASH adaptation algorithms portant features. over NDN and TCP/IP, finding NDN performance to notably exceed that of TCP/IP given certain network conditions. New adaptation algorithms specific to NDN have also been context, only a few input variables have a significant impact on proposed, such as NDNLive [6], which uses a simple RTT QoE. Note, however, that sparse Bayesian estimation is usually mechanism to stream live content with minimal rebuffering. computationally expensive. Hence, we develop here a fast new In this work, we model the video quality adaptation problem inference scheme to support online quality adaptation. as a contextual bandit problem assuming a linear parametriza- Our contributions in this paper can be summarized as: tion, which has successfully been used, e.g., for ad placement [7]. Another promising approach is based on cost-sensitive • We formulate the quality adaptation decision for QoE classification in the bandit setting [8]. Recently, [9] has maximization in ABR video streaming as a contextual discussed the use of variational inference in the bandit set- multi-armed bandit problem. ting, wherein Thompson sampling is considered to cope with • We provide a sparse Bayesian contextual bandit al- the exploration-exploitation trade-off. By assuming a high- gorithm, denoted CBA, which is computationally fast dimensional linear parametrization, we make use of sparse enough to provide real-world video players with quality estimation techniques. High-dimensional information arises adaptation decisions based on the network context. in video streaming due to the network context. Sparsity has • We show emulation testbed results and demonstrate the been a major topic in statistical modeling and many Bayesian fundamental differences to the established state-of-the-art approaches have been proposed. Traditionally, double expo- quality adaptation algorithms, especially given an NDN nential priors which correspond to ` regularization have been architecture. 1 used. However, these priors often fail due to limited flexibility 1 The developed software is provided here . The remainder in their shrinkage behavior. Other approaches that induce of this paper is organized as follows: In Sect. II, we review sparsity include ’spike-and-slab’ priors [10] and continuous relevant related work on ABR video streaming and contextual shrinkage priors. Between these two, continuous shrinkage pri- bandits. In Sect. III, we present