Search

This paper presents a detailed characterisation of user behaviour for a series of interactive video experiments over a 12 month period, in which we served popular sporting and musical content. In addition to generic VCR-like features, our custom-built Video-on-Demand application provides advanced interactivity features such as bookmarking. The dramatic impact of such functionality on how users consume content is studied and analysed. We discuss in detail how this user behaviour can be exploited by content distributors to improve user experience. Specifically, we study how simple dynamic bookmark placement and interactivity-aware content pre-fetching and replication can reduce the impact of highly interactive media on CDN performance.

Improvements at the physical network layer have enabled technologies such as 10 Gigabit Ethernet. Single core end-systems are unable to fully utilise these networks, due to limited clock cycles. Using a Multi-core architecture is one method which increases the number of available cycles, and thus allow networks to be fully utilised. However, using these systems creates a new set of challenges for network protocol processing, for example, deciding how best to utilise many cores for high network performance. This paper examines different ways the cores of a multi-core system can be utilised, and, by experimentation, we show that in an eight core system deciding which cores to use is important. In one test, there was a 40% discrepancy in CPU utilisation depending on which cores were used. This discrepancy results from the resources each core shares, an example being the multi-hierarchy CPU caches, and to which bus the processors are connected.

With the advent of numerous video distribution services, content distribution networks (CDNs) are under increasing demand. Given the associated expenses, many organisations have made use of peer-to-peer (P2P) approaches to offset bandwidth costs. Unfortunately, using clients as part of the delivery process can vastly increase load on service providers' networks. Furthermore, many approaches thus far lack the ability for interactive, on-demand viewing. For these reasons, this paper considers how 'hybrid' methods could alleviate these issues, through simulations of delivery methodologies in a CDN context. As it is unlikely that any one method will suffice for all content, network and audience types, a successful system should adapt in response to changes, as outlined in this paper.

Most existing DHT algorithms assume that all nodes have equal capabilities. This assumption has previously been shown to be untrue in real deployments, where the heterogeneity of nodes can actually have a detrimental effect upon performance. We now acknowledge that nodes on the same overlay may also differ in terms of their trustworthiness. However, implementing and enforcing security policies in a network where all nodes are treated equally is a non-trivial task. We therefore extend our previous work on Stealth DHTs to consider the differentiation of nodes based on their trustworthiness rather than their capabilities alone.

This paper presents a detailed characterisation of user behaviour for a series of interactive sport videos from the 2006 FIFA World Cup. In addition to generic VCR-like features, our custom-built Video-on-Demand architecture enabled us to provide advanced interactivity features such as bookmarking. We illustrate how such functionality may have a dramatic impact on how users consume content. A detailed discussion is also provided on how content distributors may turn this knowledge to their advantage, and thus increase the efficiency of their delivery networks.

We propose several models based on discrete-time Markov chains for the analysis of Distributed Hash Tables (DHTs). Specifically, we examine the Pastry routing protocol, as well as a Stealth DHT adaptation of Pastry to compute their exact expressions for average number of lookup hops. We show that our analytical models match with the protocols’ simulation results almost perfectly, making them ideal for rapid evaluation.

Most existing DHT algorithms assume that all nodes have equal capabilities. This assumption has previously been shown to be untrue in real deployments, where the heterogeneity of nodes can actually have a detrimental effect upon performance. In this paper, we acknowledge that nodes on the same overlay may also differ in terms of their trustworthiness. However, implementing and enforcing security policies in a network where all nodes are treated equally is a non-trivial task. We therefore extend our previous work on Stealth DHTs to consider the differentiation of nodes based on their trustworthiness rather than their capabilities alone.

The advances in wireless networking and the consequent emergence of new applications that wireless networks increasingly support inevitably leads to low capability mobile nodes connecting to peer-to-peer networks. However, the characteristics of mobile nodes and limitations of access point coverage often cause mobile nodes to lose connectivity, which can cause many mobile nodes to simultaneously rejoin the network. Continuous departures and joins due to the mobility of nodes leads to mobility churn, which can often degrade the performance of the underlying peer-topeer network significantly. In this paper, we use simulations to demonstrate that the Stealth Distributed Hash Table (Stealth DHT) algorithm is ideally suited for networks with mobile nodes. By avoiding storing state in unreliable nodes, a Stealth DHT prevents mobile nodes from being used by other nodes to provide services. Consequently, Stealth DHTs eliminate the mobility churn effect and significantly reduce the amount of overhead as compared to a generic DHT. This paper demonstrates this using simulation results that compare the performance of Stealth DHTs to a generic DHT, Pastry.

Most Distributed Hash Tables (DHTs) simply consider interconnecting homogeneous nodes on the same overlay. However, realistically nodes on a network are heterogeneous in terms of their capabilities. Because of this, traditional DHTs have been shown to exhibit poor performance in a real-world environment. Additionally, we believe that it is this approach that contributes to a limited exploitation of peer-to-peer technologies. Previous work on super-peers in DHTs was proposed to address these performance issues, however the strategy used is often based on locally clustering peers around individual super-peers. This method of superpeering, however, compromises fundamental features such as load-balancing, resilience and routing efficiency, which traditional DHTs originally promised to offer. We propose a Stealth DHT which addresses the deficiencies of previous super-peer approaches by using the DHT algorithm itself to select the most appropriate super-peer for each message sent by peers. Through simulations and measurements, we show the fitness for purpose of our proposal.

The traditional start-to-finish playback model is not suitable for all modern interactive video streams. Users expect support for higher levels of interactivity such as fast forward and rewind or the ability to arbitrary seek within their media quickly and efficiently. By conducting user studies we have observed start-to-finish is not applicable to many genres of video, and that different playback models fit better. We discuss how existing delivery techniques are impacted by these new observations. Novel interactive controls such as bookmarks have also highly impacted user behaviour. This has lead to the segments within the media being accessed in a uneven fashion, causing hotspots of interest to form; areas with orders of magnitudes more viewers than others. These hotspots typically began at the beginning of a bookmark, however not always, which lead us to design a dynamic bookmark positioning algorithm. As well as their position, determining the hotspot's length can be beneficial. This aids in autonomic techniques such as replication and pre-fetching as well as allowing the users to find what they want quicker. Under high level of interactivity, delivery techniques are less efficient due to the unpredictability of the users. We however developed techniques which restore some of this predictability, allowing clients or servers to predict future actions based on past user actions. These technique proves exceeding useful for pre-fetching which reduces seek latencies for client and can reduce load on servers. However knowledge of past user activities need to be gathered from network, thus we develop techniques to do this in a distributed manner.

The traditional start-to-finish playback model is not suitable for all modern interactive video streams. Users expect support for higher levels of interactivity such as fast forward and rewind or the ability to arbitrary seek within their media quickly and efficiently. By conducting user studies we have observed start-to-finish is not applicable to many genres of video, and that different playback models fit better. We discuss how existing delivery techniques are impacted by these new observations. Novel interactive controls such as bookmarks have also highly impacted user behaviour. This has lead to the segments within the media being accessed in a uneven fashion, causing hotspots of interest to form; areas with orders of magnitudes more viewers than others. These hotspots typically began at the beginning of a bookmark, however not always, which lead us to design a dynamic bookmark positioning algorithm. As well as their position, determining the hotspot's length can be beneficial. This aids in autonomic techniques such as replication and pre-fetching as well as allowing the users to find what they want quicker. Under high level of interactivity, delivery techniques are less efficient due to the unpredictability of the users. We however developed techniques which restore some of this predictability, allowing clients or servers to predict future actions based on past user actions. These technique proves exceeding useful for pre-fetching which reduces seek latencies for client and can reduce load on servers. However knowledge of past user activities need to be gathered from network, thus we develop techniques to do this in a distributed manner.