Performance of Peer-Assisted File Distribution

Introduction

The Internet is becoming a pervasive medium for file distribution, and growing number of users and file sizes introduce new performance challenges in the efficient distribution of files. Global Internet traffic has increased by eightfold in the last five years and is estimated to increase by another fourfold over the next five years (Cisco, 2010). File sizes have grown from several tens of megabytes, such as music files, to several gigabytes, for movies today. Traditionally, file distribution is based on the simple client-server approach that has limited scalability (Das, et al., 2006). Recently, peer-assisted approach has been proposed (Cohen, 2003; Kazza). In this approach, each peer (client) downloads a file from the server and at the same time disseminates the file to other to improve download performance.

Previous work on performance analysis of peer-assisted systems focuses on steady-state (Guo, et al., 2007; Qiu & Srikant, 2004). This is reasonable in file-sharing because peers remain in the system to continue sharing the file after completion of download. In contrast, peers in file distribution systems download the file as fast as possible and leave the system. Furthermore, the performance of peer-assisted file distribution is affected by flash crowds. In a typical file distribution system, peer arrival rate over time is characterized by two main phases as shown in Figure 1. When a new file is introduced, peer arrival surges causing flash crowd but as file popularity drops, the system goes into steady-state. Flash crowds increase user download time and thus have serious consequences on business revenue. For example, Amazon estimates that for every 100 ms delay caused by flash crowd, sales decrease by 1%, and Google reports that for each half a second increase in waiting time results in a 20% decrease in search traffic revenue (Zhang, et al., 2011). Zhang et al. (2011) show that flash crowds affect download performance for up to 45% of the users in BitTorrent (BT) (Cohen, 2003) systems. Flash crowds demand higher server bandwidth, and introduce performance challenge in provisioning the servers to deliver acceptable quality of service.

Figure 1.

Evolution of peer arrival rate in swarms

This article focuses on recent approaches in performance analysis of peer-assisted file distribution, new results for flash crowds and emerging research challenges. Firstly, basic concepts in file distribution and related works are introduced. Secondly, the article presents new insights in flash crowd performance using both measurement and analytical model, and discusses performance challenges for users, service providers and protocol designers. Lastly, this article highlights issues for achieving efficient file distribution in cellular networks where peer download capacities is lower and connectivity is less reliable.

Background

This section reviews file distribution protocols with different incentive schemes, and performance studies of peer-assisted systems at steady-state and with flash crowds.