We are witnessing a proliferation of video in the Internet; YouTube is the most bandwidth intensive service of today’s Internet. It accounts for 20 - 35% of the Internet traffic with 35 hours of videos uploaded every minute and more than 700 billion playbacks in 2010. Netflix, a web service that streams premium contents such as TV series, shows, and movies, consumes 30% of the network bandwidth in North America at peak time. Recently, leveraging the content distribution networks (CDNs), a new paradigm for video streaming on the Internet has emerged, namely, Adaptive HTTP Streaming (AHS). AHS has become the industry standard for video streaming over the Internet adopted by broadcast networks as well as VoD services such as YouTube, Netflix, Hulu, etc. In the 90’s and early 2000’s, Internet-based video streaming for high-bitrate video was challenging due to hardware limitations. In that era, to cover the hardware limitations, every software component of a video server needed to be carefully optimized to support the real-time guarantees for jitter-free video delivery. However, most of the software solutions have become less important with the remarkable hardware improvements over the past two decades. There is 100× speedup in CPU speeds; RAM capacity has increased by 1,000×; hard disk drive (HDD) capacity has grown by 10,000×. Today, CPU is no longer a bottleneck for video streaming. On the other hand, storage bandwidth and network bandwidth are still serious bottlenecks for large scale on-demand video streaming. In this dissertation, we aim at a scalable video content distribution system that addresses both storage bottleneck and network bottleneck. The first part of the dissertation pertains to the storage system on the server side: A multi-tiered storage system that exploits a flash memory solid-state drive (SSD) can meet the bandwidth needs in a much more cost- effective way than a traditional two-tier storage system. We first identify the challenges in architecting such a system given the performance quirks of flash-based SSDs, and the lim- itations of state-of-the-art multi-tiered storage systems for video streaming. Armed with the knowledge of these challenges, we show how to construct such a storage system and implement a real web server with multi-tiered storage, evaluate the system with AHS work- loads, and demonstrate significant performance gains while reducing the TCO. The second part of the dissertation pertains to the network system on the client side: Integrating peer- to-peer (P2P) technology with the client-server paradigm results in a much more scalable video content distribution system. AHS is a paradigm for client-driven video streaming; its philosophy matches well with that of P2P video streaming. An adaptation mechanism is the most important component of AHS that determines overall video streaming quality and user experience. We show a throughput-smoothing-based adaptation mechanism that is designed for a client-server architecture does not work well for a P2P architecture. We pro- vide a buffer-based adaptation mechanism, evaluate our solution with OMNeT++/INET simulator, and demonstrate significant performance gains.

In recent years, adaptive HTTP streaming protocols have become the de facto standard in the industry for the distribution of live and video-on-demand content over the Internet. In this thesis, we solve the problem of distributing adaptive HTTP live video streams to a large number of viewers using peer-to-peer (P2P) overlays. We do so by assuming that our solution must deliver a level of quality of user experience which is the same as a CDN while trying to minimize the load on the content provider’s infrastructure. Besides that, in the design of our solution, we take into consideration the realities of the HTTP streaming protocols, such as the pull-based approach and adaptive bitrate switching. The result of this work is a system which we call SmoothCache that provides CDN-quality adaptive HTTP live streaming utilizing P2P algorithms. Our experiments on a real network of thousands of consumer machines show that, besides meeting the the CDN-quality constraints, SmoothCache is able to consistently deliver up to 96% savings towards the source of the stream in a single bitrate scenario and 94% in a multi-bitrate scenario. In addition, we have conducted a number of pilot deployments in the setting of large enterprises with the same system, albeit tailored to private networks. Results with thousands of real viewers show that our platform provides an average offloading of bottlenecks in the private network of 91.5%. These achievements were made possible by advancements in multiple research areas that are also presented in this thesis. Each one of the contributions is novel with respect to the state of the art and can be applied outside of the context of our application. However, in our system they serve the purposes described below. We built a component-based event-driven framework to facilitate the development of our live streaming application. The framework allows for running the same code both in simulation and in real deployment. In order to obtain scalability of simulations and accuracy, we designed a novel flow-based bandwidth emulation model. In order to deploy our application on real networks, we have developed a network library which has the novel feature of providing on-the-fly prioritization of transfers. The library is layered over the UDP protocol and supports NAT Traversal techniques. As part of this thesis, we have also improved on the state of the art of NAT Traversal techniques resulting in higher probability of direct connectivity between peers on the Internet. Because of the presence of NATs on the Internet, discovery of new peers and collection of statistics on the overlay through peer sampling is problematic. Therefore, we created a peer sampling service which is NAT-aware and provides one order of magnitude fresher samples than existing peer sampling protocols. Finally, we designed SmoothCache as a peer-assisted live streaming system based on a distributed caching abstraction. In SmoothCache, peers retrieve video fragments from the P2P overlay as quickly as possible or fall back to the source of the stream to keep the timeliness of the delivery. In order to produce savings, the caching system strives to fill up the local cache of the peers ahead of playback by prefetching content. Fragments are efficiently distributed by a self-organizing overlay network that takes into account many factors such as upload bandwidth capacity, connectivity constraints, performance history and the currently being watched bitrate. / <p>QC 20131122</p>