Modern video players employ complex algorithms to adapt the bitrate of the video that is shown to the user. Bitrate adaptation requires a tradeoff between reducing the probability that the video freezes and enhancing the quality of the video shown to the user. A bitrate that is too high leads to frequent video freezes (i.e., rebuffering), while a bitrate that is too low leads to poor video quality. Video providers segment the video into short chunks and encode each chunk at multiple bitrates. The video player adaptively chooses the bitrate of each chunk that is downloaded, possibly choosing different bitrates for successive chunks. While bitrate adaptation holds the key to a good quality of experience for the user, current video players use ad-hoc algorithms that are poorly understood. We formulate bitrate adaptation as a utility maximization problem and devise an online control algorithm called BOLA that uses Lyapunov optimization techniques to minimize rebuffering and maximize video quality. We prove that BOLA achieves a time-average utility that is within an additive term O(1/V) of the optimal value, for a control parameter V related to the video buffer size. Further, unlike prior work, our algorithm does not require any prediction of available network bandwidth. We empirically validate our algorithm in a simulated network environment using an extensive collection of network traces. We show that our algorithm achieves near-optimal utility and in many cases significantly higher utility than current state-of-the-art algorithms. Our work has immediate impact on real-world video players and for the evolving DASH standard for video transmission.

Using more than 12,000 servers in over 1,000 networks, Akamai's distributed content delivery system fights service bottlenecks and shutdowns by delivering content from the Internet's edge. As Web sites become popular, they're increasingly vulnerable to the flash crowd problem, in which request load overwhelms some aspect of the site's infrastructure, such as the frontend Web server, network equipment, or bandwidth, or (in more advanced sites) the back-end transaction-processing infrastructure. The resulting overload can crash a site or cause unusually high response times -both of which can translate into lost revenue or negative customer attitudes toward a product or brand.Our company, Akamai Technologies, evolved out of an MIT research effort aimed at solving the flash crowd problem (www.akamai.com/en/html/about/history. html). Our approach is based on the observation that serving Web content from a single location can present serious problems for site scalability, reliability, and performance. We thus devised a system to serve requests from a variable number of surrogate origin servers at the network edge.1 By caching content at the Internet's edge, we reduce demand on the site's infrastructure and provide faster service for users, whose content comes from nearby servers.When we launched the Akamai system in early 1999, it initially delivered only Web objects (images and documents). It has since evolved to distribute dynamically generated pages and even applications to the network's edge, providing customers with on-demand bandwidth and computing capacity. This reduces content providers' infrastructure requirements, and lets them deploy or expand services more quickly and easily. Our current system has more than 12,000 servers in over 1,000 networks. Operating servers in many locations poses many technical challenges, including how to direct user requests to appropriate servers, how to handle failures, how to monitor and control the servers, and how to update software across the sys- Global Deployment of Data Centerstem. Here, we describe our system and how we've managed these challenges. Existing ApproachesResearchers have explored several approaches to delivering content in a scalable and reliable way. Local clustering can improve fault-tolerance and scalability. If the data center or the ISP providing connectivity fails, however, the entire cluster is inaccessible to users. To solve this problem, sites can offer mirroring (deploying clusters in a few locations) and multihoming (using multiple ISPs to connect to the Internet). Clustering, mirroring, and multihoming are common approaches for sites with stringent reliability and scalability needs. These methods do not solve all connectivity problems, however, and they do introduce new ones:I It is difficult to scale clusters to thousands of servers. I With multihoming, the underlying network protocols -in particular the border gateway protocol (BGP) 2 -do not converge quickly to new routes when connections fail. I Mirroring requires synchronizing the site amo...

The distribution of videos over the Internet is drastically transforming how media is consumed and monetized. Content providers, such as media outlets and video subscription services, would like to ensure that their videos do not fail, start up quickly, and play without interruptions. In return for their investment in video stream quality, content providers expect less viewer abandonment, more viewer engagement, and a greater fraction of repeat viewers, resulting in greater revenues. The key question for a content provider or a content delivery network (CDN) is whether and to what extent changes in video quality can cause changes in viewer behavior. Our work is the first to establish a causal relationship between video quality and viewer behavior, taking a step beyond purely correlational studies. To establish causality, we use Quasi-Experimental Designs, a novel technique adapted from the medical and social sciences. We study the impact of video stream quality on viewer behavior in a scientific data-driven manner by using extensive traces from Akamai's streaming network that include 23 million views from 6.7 million unique viewers. We show that viewers start to abandon a video if it takes more than 2 s to start up, with each incremental delay of 1 s resulting in a 5.8% increase in the abandonment rate. Furthermore, we show that a moderate amount of interruptions can decrease the average play time of a viewer by a significant amount. A viewer who experiences a rebuffer delay equal to 1% of the video duration plays 5% less of the video in comparison to a similar viewer who experienced no rebuffering. Finally, we show that a viewer who experienced failure is 2.32% less likely to revisit the same site within a week than a similar viewer who did not experience a failure.Index Terms-Causal inference, Internet content delivery, multimedia, quasi-experimental design, streaming video, user behavior, video quality.

in part by an NSF CAREER Award CCR-9703017. 1 We use with high probability to mean with probability at least 1 ; O(1=n) for some constant generally this will be 1. A precise analysis shows that the expected maximum load is ; ;1 (n) ; 3=2 + o(1) Gon81].

No abstract