LJMU Research Online

Yang, P, Fan, Q, Yin, H, Min, G, Luo, Y, Lyu, Y, Huang, H and Jiao, L

Video Delivery Networks: Challenges, Solutions and Future Directions http://researchonline.ljmu.ac.uk/id/eprint/6324/

Article

Citation (please note it is advisable to refer to the publisher’s version if you intend to cite from this work)

Yang, P, Fan, Q, Yin, H, Min, G, Luo, Y, Lyu, Y, Huang, H and Jiao, L (2017) Delivery Networks: Challenges, Solutions and Future Directions. Computers and Electrical Engineering, 66. pp. 332-341. ISSN 0045-7906

LJMU has developed LJMU Research Online for users to access the research output of the University more effectively. Copyright © and Moral Rights for the papers on this site are retained by the individual authors and/or other copyright owners. Users may download and/or print one copy of any article(s) in LJMU Research Online to facilitate their private study or for non-commercial research. You may not engage in further distribution of the material or use it for any profit-making activities or any commercial gain.

The version presented here may differ from the published version or from the version of the record. Please see the repository URL above for details on accessing the published version and note that access may require a subscription.

For more information please contact [email protected]

http://researchonline.ljmu.ac.uk/ Video Delivery Networks: Challenges, Solutions and Future Directions

Qilin Fan1, Hao Yin1, Geyong Min2, Po Yang3, Yan Luo4, Yongqiang Lyu1, Haojun Huang5 and Libo Jiao1

1Research Institute of Information Technology (RIIT), Tsinghua University 2Department of Mathematics and Computer Science, University of Exeter 3Department of Computer Science, Liverpool John Moores University 4Department of Electrical and Computer Engineering, University of Massachusetts Lowell 5Department of Communication Engineering, Wuhan University

ABSTRACT ing higher definition video streams, requesting more and ecosystems are faced with the increasing re- more bandwidth. The Cisco predicts that video con- quirements in versatile applications, ubiquitous consump- sumption will amount to 82% of the global consumer tion and freedom of creation and sharing, in which the user trac in 2020 [1]. In conjunction with these growing experience for high-quality services has become more and trac volumes, users’ expectations of high quality of more important. Internet is also under tremendous pres- experience (e.g., high-definition video, low re-bu↵ering, sure due to the exponential growth in video consumption. low startup delays) are continuously increasing [2]. Video providers have been using content delivery networks Content Delivery Networks (CDNs) have been playing (CDNs) to deliver high-quality video services. However, the a critical role in content hosting, orchestration, mashup, new features in video generation and consumption require transport, and edge access. Today, video providers rely CDN to address the scalability, quality of service and flex- on CDNs, such as Akamai and Limelight to serve video ibility challenges. As a result, we need to rethink future content across di↵erent geographical locations. However, CDN for sustainable video delivery. To this end, we give with the new characteristics of Internet video generation an overview for the Internet video ecosystem evolution. We and consumption, the architecture of today’s HTTP- survey the existing video delivery solutions from the per- based CDN is being stressed [3]. spective of economic relationships, algorithms, mechanisms As a result, there is a strong call for rethinking future and architectures. At the end of the article, we propose a evolution for sustainable video delivery. The state of data-driven information plane for video delivery network as the art in video delivery network has experienced a the future direction and discuss two case studies to demon- major boost in recent years, measured in the number of strate its necessity. publications and industrial groups focused on the topic. Diverse interests have led to heterogeneous, fragmented Keywords landscapes. In this article, we give a comprehensive review and analysis on video delivery network, focusing Video Delivery Network, Scalability, Quality of Ser- on the major obstacles in the way of further progress vice, Cloud computing, Data-Driven, Information Plane and suggesting possible future direction. The rest of the article is organized as follows. We 1. INTRODUCTION firstly give an overview of both the evolution of Inter- Video is a kind of important media for communication net video ecosystem and the key challenges to video and entertainment. In the past decade, the consump- delivery network in new personalization phase. After a tion of video contents changed from o✏ine viewing to brief summary on the surveys in this field, we illustrate online streaming, which imposed stringent requirements how existing approaches work to address parts of chal- on the network latency and throughput of video delivery lenges from the perspective of economic relationships, network. Internet video ecosystem has experienced new algorithms, mechanisms and architectures. Finally, we characteristics accordingly: (1) the diversity of video suggest constructing data-driven multidimensional in- applications, such as online streaming (e.g., IPTV and formation plane for video delivery network as future Netflix), and real-time video conferencing (e.g., Skype), research trend and conclude the article. (2) the ubiquitous video consumption by users who watch wherever they are and whenever they can; and (3) the freedom in the creating, distributing and sharing 2. VIDEO DELIVERY NETWORK: EVOLU- of user generated video contents through online video TION AND CHALLENGES services such as YouTube, Vimeo or Hulu. The situation Figure 1 shows the evolution of video ecosystem expe- is further deteriorated by the emerging trends of adopt- rienced several important phases in the last two decades. bu↵ering ratio reduces user engagement by 3 minutes in 1.#Streaming## 2.#Web#Video# 3.#Personaliza9on# Phase# Phase Phase# 2010 but 6 minutes in 2012. Therefore, video delivery !Simple(video(playback( -Integrated(with(Web(browser( !UGC,(SNS(consumpKon( service providers must keep improving server-side and !Proprietary(protocols( !RTMP(protocol( !MulKple(devices(and(apps( !Streaming(technology( !CDN,(P2P(technology(( !HTTP(streaming(protocol( network-level performance for better QoS. !Cloud(compuKng,(SDN,((((( data!driven,(mobile…( 3) High flexibility: Video consumption has shifted technology((( ( dramatically from a relatively predictive and controlled model to a highly dynamic mode, such as the large trac spikes during online live and on-demand events. Streaming large live videos and popular on-demand ti- tles can cause significant network congestion leading to 1995 2004 2005 2011 Timeline poor subscriber experience. Traditional CDNs, however, are statically deployed, and their centralized redirection Figure 1: Evolution of Internet video ecosystem. mechanisms are not good for dynamic adjustment, which results in performance penalties. On the other hand, traditional CDNs were horizontal - designed to deliver In the mid-1990s, the Internet boosted from a text-based all forms of content by the same mechanism across all system to a multimedia-contained system. The repre- user categories. The flexible service depending on di↵er- sentative streaming technique changed the way videos ent business models require new CDNs that should be that are transmitted over the Internet, from the time- vertically customized according to di↵erent content and consuming downloads to real-time viewing. In 2005, user requirements. Adobe system acquired the flash player which was origi- nally developed by Macromedia, and flash soon became the de-facto standard for web-based streaming video 3. EXISTING SOLUTIONS (over RTMP) in this wave when videos explosively trans- To address the aforementioned challenges, a range of mitted across the Internet along with widespread broad- approaches were proposed in recent years. We survey the band access. Meanwhile, CDN and Peer-to-Peer (P2P) existing methods and classify them into four categories technologies emerged with the tide of the times. according to the challenge type the approach focused Since 2011, the portable devices with cameras have on and the perspective from which the approach solved become more and more ubiquitous, which enable users (shown in table 1). conveniently create and share videos. Users are increas- ingly demanding content on their terms - any content, 3.1 Cooperation between stakeholders any time, any device, any place. In addition, another The stakeholders in video delivery network ecosystem key driver to the rapid explosion of streaming video was can be grouped into the client, CDN, Internet Service the technical shift from specialized streaming protocols Provider (ISP), and Content Provider (CP). CP has a such as RTMP to HTTP chunk-based streaming proto- central database of content and pays CDN for content cols. The use of commodity service greatly decreases delivery service. The CDN is an overlay network that the barrier of access entry by allowing video providers to o↵ers additional value to the basic Internet content utilize the existing HTTP-based CDN infrastructures in delivery. It pays for transit service provided by ISP. delivering videos to wide audiences. New applications, Client pays ISP for network access and pays CP for devices and protocols not only promote the Internet watching videos as well. Traditional CDNs do not always video ecosystem in the personalization phase, but also have enough service points in locations that can provide bring new challenges to the video delivery network. The good performance to satisfy the growing demands. To key challenges can be summarized as follows. address such a scalability issue, the boundaries among 1) High scalability: We are witnessing a prolifera- di↵erent stakeholders in video delivery networks are tion of videos over the Internet. For example, Netflix, blurring for cooperation. An overview of the involvement a web service that streams premium video contents, ac- of content delivery stakeholders is shown in figure 2. counted for 37 percent of peak download Internet trac Driven by economic policies, several new hybrid ar- in North America in 2015. As of July 2015, more than chitectures were proposed: 400 hours of video were uploaded to YouTube every 1) CPs deploy application-specific CDNs with direct minute, up from 300 hours per minute in November peering with or inside ISPs, e.g., Netflix’s platform - 2014. According to Cisco report [1], IP video trac Open Connect for video stream delivery and will be 82% of all consumer Internet trac by 2020, up Global Cache, primarily for YouTube. It allows ISPs to from 70% in 2015. Today, boosted by the strong growth cache the content locally, and thus reduces the strain of of newly uploaded videos and demand for various bi- CP’s trac over the network. trate, storage and network bandwidth become serious 2) Commercial CDNs and clients operate hybrid delivery bottlenecks for large scale video delivery. systems where clients download contents from servers 2) High Quality of Service (QoS): Video contents as well as from other clients. CDNs provide excellent are being increasingly consumed on big screens. quality when workloads are within the provisioning lim- Users expect much higher quality of video service than its; otherwise, P2P technologies can be used to solve ever. The existing study [2] showed that 1% increase in the scalability issue by utilizing the resources of the Table 1: The categories of the existing solutions. What does it focus on? From which perspective? How does it solve? Scalability Economic relationship adjustment Cooperation between stakeholders (Sec. 3.1) QoS Algorithm optimization Bitrate and CDN/server adaption (Sec. 3.2) Flexibility Infrastructure as a service Cloud computing utilization (Sec. 3.3) Scalability Architecture redesign Information-centric network (Sec. 3.4)

,,,,CDNI, three dimensions [3]: 1) What parameters can we control? Bitrate or CDN/server; 2) Who decides the values for tuned parameters? Client-

CP CDN ISP Client side, server-driven or control plane; 3) When can we choose the parameters? Startup or midstream.

,,,Applica1on,CDN, ,,,ISP3operated,CDN, Table 2 illustrates several representative work in this field by combining di↵erent options from these three ,,Hybrid,CDN3P2P, dimensions. Clients are very sensitive to local network dynamics. Figure 2: Involvement of content delivery stake- At the same time, they cannot respond eciently to holders. the network variations due to their distributed nature. Therefore, most bitrate adaptive algorithms are driven participating peers. This hybrid CDN-P2P architecture by client-side [7][8][9][10]. Nevertheless, the advantage has been widely used in live streaming and social video of server-side bitrate adaptive algorithms lies in not re- propagation[4]. quiring clients’ cooperation and solving the problem of 3) ISPs, on one hand, are vulnerable to the tracdy- unfairness and instability [11][12]. There are also some namics caused by time-varying shifts in content demand. other studies that uncover proprietary CDN/server se- On the other hand, they wish to take advantage of their lection algorithms [13][14]. A client in Netflix adapts regional presence and proximity to end users to enter to bandwidth variations by adjusting the bitrates and the content delivery market [5]. This leads many ISPs to continues to use the same CDN server only when the deploy proprietary CDNs within their networks, which current CDN server is unable to serve the lowest bitrate, allows them to obtain extra income from CP by o↵ering it switches to a di↵erent CDN server. Inspired by in- their CDN services to them as a complement to pure- dustrial CDN multihoming applications, Liu et al. [15] play CDNs. optimized the CDN selection to make a client actively 4) A further-reaching solution was also discussed when use additional servers when the service of a single CDN Cisco began to explore the feasibility of CDN federa- server is insucient. Study [3][16] even designed a video tions since 2011. CDNs in federation are supposed to control plane that can use a global view of CDN and net- be interconnected via open interfaces and shared by au- work performance to dynamically assign clients suitable tonomous members so they can act as a multi-footprint CDNs and bitrates. logical CDN. Internet Engineering Task Force (IETF) The motivation of bitrate adaptive algorithms or working group was also established to define the stan- CDN/server selection is to deliver as high quality video dards of CDN interconnection (CDNI) [6]. as possible even in dynamic environment. However, the The alliances between stakeholders show a natural algorithms use measurement-driven performance feed- evolution of video delivery network today, which was back without knowledge of bottleneck from the network driven by the requirements for storage and bandwidth level. Therefore, there appears to be some more space issues. However, such alliances require meticulous mech- for performance improvement with network-level infor- anism design and management to guarantee the fairness mation from ISP given. among the stakeholders. 3.3 Cloud Computing Utilization 3.2 Bitrate and CDN/Sever Adaption To meet volatile user demands for videos, cloud com- The research on the impact of service quality [2] puting emerges as a new paradigm to reshape the pattern showed that users are quite sensitive to bu↵ering and of video distribution over the Internet, in which resources high startup latency and they prefer to higher bitrate (such as CPU, bandwidth, and storage) can be rented videos. It also identifies that many of the user experi- on demand from cloud data centers. In cloud comput- ence issues today are caused by sub-optimal client-side ing environment, hardware and software resources are control logic, and spatial and temporal diversities of abstracted using virtualization technology for providing performance over di↵erent CDNs and networks. service to consumers. The design space of adaptive algorithms for video Industrial practitioners have made some e↵orts to delivery quality in recent work can be decomposed into migrate VoD applications to the cloud. Netflix, a major Table 2: Representative work on adaptive algorithms for video delivery quality. What parameter? Who choose? When to choose? Related work Bitrate Client Midstream [7][8][9][10] Bitrate Server Midstream [11][12] CDN/Server Control plane Startup [13] CDN/Server Control plane + Client Midstream [15] Bitrate Client Midstream [14] CDN/Server Control plane Midstream Bitrate Control plane Midstream [3][16] CDN/Server Control plane Midstream

VoD provider in North America, has moved its streaming 3.4 Information-Centric Networking servers, encoding software, search engines, and huge Since the current growth rate for the bandwidth de- data stores to Amazon Web Services (AWS). Netflix’s mand of video applications is dicult to sustain for successful practice verifies that cloud computing service the network operators, new architecture - information- could provide elastic resources even for large-scale video centric networking (ICN) has been proposed. It allows service. the network to cache data by attaching storage to routers, There have also been many studies on utilizing cloud eliminating complex optimizations required by CDNs, computing resources in academic circles: so that content can be distributed in a cost-ecient and 1) With intensive and dynamic bandwidth/storage de- secure manner. This is rooted by the observation that mands in real time, one fundamental question is: how Internet trac consumption is largely content driven. to quantify dynamic user demands in service delivery, Recent studies extended ICN to support video stream- or more precisely, how should a video service provider ing. Video streaming over ICN has faced two challenges learn the dynamic demands from users? Wu et al. [17] with respect to performance. First, Adaptive Bitrate introduced a Jackson queueing network based model Streaming (ABS) is incompatible with ICN’s funda- to characterize the viewing behaviors of users inside mental principles. For instance, an ABS connection each channel in the VoD system and studied the upload does not support swarming which is natural in ICN. capacity provided by cloud resources in C/S and P2P The situation becomes more complicated when multiple mode. users requesting the same content but with di↵erent 2) Video delivery service is a network-bound service with access bandwidth. Yu et al. [20] proposed a DASH- stringent bandwidth requirements. Therefore, band- aware scheduling algorithm that the network (by way of width reservation becomes a value-added feature o↵ered logically centralized controller) computes the available by cloud services to attract customers with bandwidth- link resources and schedules the chunk dissemination to intensive applications. It can be translated into the edge caches ahead of the user’s requests by utilizing the question: how should a video service provider dynami- DASH manifest. Second, video trac volume inflates cally book bandwidth resources from multiple locations caching resource requirements. Therefore, caching redun- and direct user requests to guarantee user’s experience? dancy in in-network level a↵ects more severely real-time He et al. [18] applied the Nash bargaining solution to video transfers. Sun et al. [21] conducted large-scale solve the joint optimization problem that takes both the trace-driven evaluation and analysis of ICN-on-video operational cost and user experience into account and and observed that the benefits of ICN for video work- derived the optimal bandwidth provisioning strategy. loads with caches smaller than 100GB are marginal and 3) In addition to the bandwidth requirement, the stor- discussed that ubiquitous caching is not fundamentally age demand is also significant, especially for UGC or necessary. Therefore, it suggests caching only in spe- social video services which easily ignite viral propagation cific locations along the content delivery path or using e↵ects. Federation of geo-distributed cloud services pro- cooperative caching to reduce redundancy. vides a more economic motivation: “Infinite” on demand Although the content-driven principle in ICN is best resources. Its realize presents challenge on: how to e- suit for bandwidth-intensive video applications, from the ciently store and migrate video contents across di↵erent ISP’s point of view, ICN clean slate approaches require cloud sites? Wu et al. [19] designed an online algorithm intrusive modifications to the communication stack and that can utilize cloud storage resources to accommo- have yet to demonstrate trac engineering capabilities. date dynamic demand on the fly and further achieve the Moreover, addressing, naming and routing issues are not optimality via a (t)-step look ahead mechanism. yet solved to scale and need more research. Moving the whole video delivery to the cloud can be a real challenge as cloud will not replace all tradi- tional hosting or on-premise deployments but comple- 4. DATA-DRIVEN INFORMATION PLANE ment them. As a result, there will always be situations Based on the review of existing approaches, we have where security requirements, flexibility, performance or a further inspection on the root causes of the challenges. control will preclude the cloud. Current layered architecture isolates information ex- • Bitrate' • Buffering'ra,o' User' • Join',me' Experience' • …'

• Popularity' Content'Access' • LifeCcycle' Characteris,cs' • Spa,al'propaga,on' • …' • Latency' • Throughput' Network'Performance' • JiFer' • …' • Network'topology' • AS'rela,onship' Network'Infrastructure' • Complex'sta,s,cal'feature'' • …'

Figure 3: Information plane for video delivery network. Figure 4: Correlation between average delay and average deployment cost at di↵erent time. changes between video applications and networks. For video applications, the network states are invisible and network resources are uncontrollable. This also hap- We describe our servers placement framework - Netclust pens as network is unaware of application content and [24] as example. Traditional servers placement strategy user profile. Therefore, we would like to introduce a is a classical facility location problem with the aim of data-driven information plane for intermediate layer - choosing M replicas among N potential sites. Known video delivery network as a possible future direction. as a NP-hard problem, it does not scale well with the The data-driven information plane consists of four layers number of clients and the number of server candidates from bottom up in figure 3. They can be divided into two with the rapid growth of cloud computing platforms. parts. Network infrastructure and network performance It also fails in discovering the inherent rule and poten- is related to network providers. Network infrastruc- tial trends of the service performance as the number of ture layer contains network structural knowledge such servers to be deployed changes, and thus cannot give as topology, AS relationships and complex statistical insights on the choice of an ideal number of servers. features. Network performance layer reflects network Given large-scale network performance in information quality of service such as latency, throughput and jitter, plane, Netclust could leverage a clustering algorithm to which is built upon network infrastructure. Content determine the correlation between the average service access characteristics and user experience is related to delay and the unit deployment cost for busy period and content providers. Content access characteristics like idle period. Figure 4 shows such relationship. Firstly, we popularity, life-cycle and spatial propagation shows spe- observe that the service delay decreases as deployment cific user behaviors in di↵erent content providers. User cost increases, following the similar negative exponential experience layer reflects quality of service provided by distribution. Secondly, the delay-cost curve of “busy content providers such as bitrate, bu↵ering ratio and join network” is converging faster than “idle network”. More- time. Based on this information plane, video delivery over, with the same deployment cost, “idle network” has network can be significantly improved via sophisticated the better performance than “busy network”. Operators processing such as control, fault diagnosis, scheduling, can easily make the most suitable decision based on the and decision making. We will take network performance relationship, analyzed above, between the number of and content access characteristics below as two case servers, deployment cost and service performance. studies to demonstrate how information plane plays an important role. 4.2 Content Access Characteristics With regard to replication and caching mechanisms 4.1 Network Performance in video delivery networks, it is natural to answer which An up-to-date annotated map of the network per- videos should be replicated, how many replicas will be, formance is significant for network management, task and where they will reside. Take for scheduling, and resource optimization in video delivery user generated content as an example, the approaches network. Progress has been made in data models and can be divided into two classes: (1) Proactive approach: tools for at-scale network measurements using minimum traditional replica placement, together with request rout- deployment cost. A network coordinate system [22] mod- ing is formulated as an optimization problem given global els the Internet as a geometric space and characterizes prior information such as contents popularity and net- the position of any node in the Internet by a coordinate work congestion. However, its high computing overhead in this space. IPlane [23] utilizes the Internets routing and lack of quick response to access interest changes topology to build an atlas of the Internet, and estimates limit its practical application. (2) Reactive approach: the properties of the paths between arbitrary end hosts the contents for caching are driven by users’ requests. on the Internet based on the path composition technique. However, its video popularity blindness characteristic

1.0 0.9 0.8 0.7 0.6 0.5

0.4 0.3 [0,1E4) [1E4,1E5) 0.2 [1E5,1E6) ProportionViews of 0.1 [1E6,∞) 0.0 147101316192225283134 Provinces Rank

Figure 5: Cumulative distribution of videos’ Figure 6: Geographic cumulative distribution of PPMCC for di↵erent temporal lag. views for videos with di↵erent popularity. leads to cache pollution, where the cache is wasted by approach to model the metric interdependencies and low demand videos, resulting in low hit rate and high their complex relationships to user engagement, aiming bandwidth cost. We have worked on understanding the at developing a predictive model of user QoE in viewing properties of videos and how the users access behaviours video. This model could guide CDN and bitrates selec- can be modelled and predicted in our recent research. tion to achieve improvement in overall user engagement. We plot cumulative distribution function of pearson The case studies discussed above demonstrate that the product-moment correlation coecient (PPMCC) for next stage of the development of video delivery network di↵erent temporal lags scaling from 1 to 7 for sampled highly relys on multidimensional data-driven informa- videos in figure 5. We observed that the median PPMCC tion plane and correlations learned from data can be for videos of temporal lag 1 is 0.78, which shows strong utilized to e↵ectively improve network resource alloca- correlation from temporal perspective. For temporal tion on demand, enhance user experience, and reduce lags scaling from 2 to 7, the median PPMCCs are still cost. larger than 0.5 but decrease slowly as the temporal lag grows. It suggests that videos popularity is predictable 5. ACKNOWLEDGEMENT using recent access times and recent 4 days values are This work was supported in part by the National Key enough for predicting. We also examined the geographic Research and Development Program under Grant no. distribution of views for videos in province granularity. 2016YFB1000102, in part by the National Natural Sci- The interval in each colored line denotes video access ence Foundation of China under Grant nos. 61672318, in times. Point (x,y) in the figure 6 means that the x top part by the projects of Tsinghua National Laboratory for locations for videos account for y% of the total num- Information Science and Technology (TNList), in part ber of views in each category. It illustrates that 30% by the European Seventh Framework Programme (FP7) provinces could cover 70% of video views for all cate- under Grant no. PIRSES-GA-2012-318939, in part by gories. This geographic locality feature illustrates that Natural Science Foundation of China no. 61402343 ad in content replicating in “appropriate small-scale locations part by Natural Science Foundation of Jiangsu/Suzhou can achieve most of the trac. Research [4] also ex- no. BK20160385. tract the content access characteristics of social videos from correlation analysis at information plane to guide propagation-aware replication and improve video service 6. CONCLUSION quality. The video delivery network in new era is a hot and active research area. In this article, we review the evolu- 4.3 Complementary tion of Internet video ecosystem and summarize the key There are also several works in the recent literature challenges video delivery network are faced with. There dealing with network infrastructure aware or user ex- are many e↵orts invested in this field, including the co- perience aware video delivery network. Studies [25][26] operation between stakeholders, optimization algorithms developed algorithms to heuristically infer AS relation- for video delivery quality, the introduction of cloud com- ships based on AS path information available in public puting and the invention of ICN architecture to partially BGP data. Accurate and complete AS relationships address scalability, QoS and flexibility challenges. In or- could provide a scientific basis for decision-making of der to fundamentally solve these challenges, we highlight cost-e↵ective cache deployment for CDN [27]. Balachan- the necessity of data-driven multidimensional informa- dran et al. [28] collected statistics directly from the tion plane for video delivery network. Along with the video player in client side and presented a data-driven case studies, we demonstrate significant practical ben- efits that video delivery network can obtain by using delivery. In Proc. of INFOCOM, pages 1620–1628, information plane. 2012. [15] H. Liu, Y. Wang, Y. Yang, H. Wang, and C. Tian. 7. REFERENCES Optimizing cost and performance for content multihoming. In Proc. of SICCOMM, pages [1] Cisco. Cisco global cloud index: Forecast and 371–382, 2012. methodology, 2015-2020. CISCO White paper, [16] A. Ganjam, J. Zhan, X. Liu, F. Siddiqi, J. Jiang, pages 1–28, 2016. V. Sekar, I. Stoica, and H. Zhang. C3: [2] F. Dobrian, A. Awan, D. Joseph, A. Ganjam, Internet-scale control plane for video quality J. Zhan, V. Sekar, I. Stoica, and H. Zhang. optimization. In Proc. of NSDI, 2015. Understanding the impact of video quality on user [17] Y. Wu, C. Wu, B. Li, X. Qiu, and F. Lau. engagement. ACM SIGCOMM CCR, 41(4), 2013. Cloudmedia: When cloud on demand meets video [3] X. Liu, F. Dobrian, H. Milner, J. Jiang, V. Sekar, on demand. In Proc. of ICDCS, pages 268–277, I. Stoica, and H. Zhang. A case for a coordinated 2011. internet video control plane. In Proc. of [18] J. He, D. Wu, Y. Zeng, X. Hei, and Y. Wen. SIGCOMM, pages 359–370, 2012. Toward optimal deployment of cloud-assisted video [4] Z. Wang, L. Sun, X. Chen, W. Zhu, J. Liu, distribution services. IEEE Transactions on M. Chen, and S. Yang. Propagation-based Circuits and Systems For Video Technology, social-aware replication for social video contents. 23(10):1717–1728, 2013. In Proc. of Multimedia, pages 29–38, 2012. [19] Y. Wu, C. Wu, B. Li, L. Zhang, Z. Li, and F. Lau. [5] Pantelis A Frangoudis, Louiza Yala, Adlen Scaling social media applications into Ksentini, and Tarik Taleb. An architecture for geo-distributed clouds. In Proc. of INFOCOM, on-demand service deployment over a telco cdn. In pages 684–692, 2012. Communications (ICC), 2016 IEEE International [20] Y. T. Yu, F. Bronzino, R. Fan, C. Westphal, and Conference on, pages 1–6. IEEE, 2016. M. Gerla. Congestion-aware edge caching for [6] Larry Peterson, Bruce Davie, and Ray van adaptive video streaming in information-centric Brandenburg. Framework for content distribution networks. In Proc. of CCNC, pages 588–596, 2015. network interconnection (cdni). Technical report, [21] Y. Sun, S. Fayaz, Y. Guo, V. Sekar, Y. Jin, 2014. M. Kaafar, and S. Uhlig. Trace-driven analysis of [7] J. Jiang, V. Sekar, and H. Zhang. Improving icn caching algorithms on video-on-demand fairness, eciency, and stability in http-based workloads. In Proc. of CoNext, pages 363–376, adaptive video streaming with festive. In Proc. of 2014. CoNext, pages 97–108, 2012. [22] T. E. Ng and H. Zhang. Global network [8] T. Huang, R. Johari, N. Mckeown, M. Trunnell, positioning: a new approach to network distance and M. Watson. A bu↵er-based approach to rate prediction. Computer Communication Review, adaptation (evidence from a large video streaming 32(1):61, 2002. service). In Proc. of SIGCOMM, pages 187–198, [23] Harsha V Madhyastha, Ethan Katz-Bassett, 2014. Thomas E Anderson, Arvind Krishnamurthy, and [9] X. Yin, A. Jindal, V. Sekar, and B. Sinopoli. A Arun Venkataramani. iplane nano: Path prediction control-theoretic approach for dynamic adaptive for peer-to-peer applications. In NSDI, volume 9, video streaming over http. In Proc. of SIGCOMM, pages 137–152, 2009. pages 325–338, 2015. [24] H. Yin, X. Zhang, T. Zhan, Y. Zhang, G. Min, and [10] S. Akhshabi, L. Anantakrishnan, A. Begen, and D. O. Wu. Netclust: A framework for scalable and C. Dovrolis. What happens when http adaptive pareto-optimal media server placement. streaming players compete for bandwidth? In Proc. Transactions on Multimedia, 15(8):2114–2124, of NOSSDAV, pages 9–14, 2012. 2013. [11] S. Akhshabi, L. Anantakrishnan, C. Dovrolis, and [25] Matthew Luckie, Bradley Hu↵aker, Amogh A. Begen. Server-based trac shaping for Dhamdhere, Vasileios Giotsas, et al. As stabilizing oscillating adaptive streaming players. relationships, customer cones, and validation. In In Proc. of NOSSDAV, pages 19–24, 2013. Proceedings of the 2013 conference on Internet [12] L. De Cicco, S. Mascolo, and V. Palmisano. measurement conference, pages 243–256. ACM, Feedback control for adaptive live video streaming. 2013. In Proc. of Multimedia Systems, pages 145–156, [26] Vasileios Giotsas, Matthew Luckie, Bradley 2011. Hu↵aker, et al. Inferring complex as relationships. [13] V. Adhikari, S. Jain, and Z. Zhang. Where do you In Proceedings of the 2014 Conference on Internet ”tube”? uncovering server selection Measurement Conference, pages 23–30. ACM, 2014. strategy. In Proc. of ICCCN, pages 1–6, 2011. [27] Syed Hasan, Sergey Gorinsky, Constantine [14] V. Adhikari, Y. Guo, F. Hao, M. Varvello, V. Hilt, Dovrolis, and Ramesh K Sitaraman. Trade-o↵sin M. Steiner, and Z. Zhang. Unreeling netflix: optimizing the cache deployments of cdns. In IEEE Understanding and improving multi-cdn movie INFOCOM 2014-IEEE Conference on Computer Communications, pages 460–468. IEEE, 2014. [28] A. Balachandran, V. Sekar, A. Akella, Seshan S., I. Stoica, and H. Zhang. Developing a predictive model of quality of experience for internet video. ACM SIGCOMM CCR, 43(4):339–350, 2013.