P2P as a CDN: A new service model for file sharing

Abstract

Traditional content distribution networks (CDNs), such as Akamai, replicate content at thousands of servers worldwide in an attempt to bring it closer to end users. Recent years have, however, brought a surge of peer-to-peer (p2p) systems that have demonstrated the ability both to help traditional CDN operations and to effectively disseminate content as independent applications. Unfortunately, this p2p surge has created significant problems to network operators by generating large volumes of inter-AS traffic. In this paper, we demonstrate that stepping back and applying traditional CDN design principles to p2p systems can help solve these emerging problems. In particular, focusing on the BitTorrent swarming protocol, we show that our new service model can, in the common case, reduce inter-AS traffic by 45–75%. Moreover, in scenarios when ISPs are shaping inter-AS traffic, it speeds up download times by 60% for the most popular torrents.

Our approach bases on disproving the common wisdom that the current peer altruism in p2p systems (BitTorrent in particular) is insufficient. We thus abandon the common approach of deploying novel incentives for cooperation, and focus on designing methods to effectively utilize existing system resources. We show that controlled regional-based content replication, common for the traditional CDN design, can effectively achieve this goal. We implement our system and demonstrate that it effectively scales. Moreover, it is incrementally deployable and brings significant benefits in partial deployment scenarios. ISPs and network regions in which the system gets deployed can resolve their inter-AS traffic problems instantly and autonomously, i.e., without any inter-ISP collaboration and without requiring that the system gets deployed in the entire Internet.

Introduction

Peer-to-peer (p2p) applications gained enormous popularity in recent years, generating a large fraction of wide-area Internet traffic [1], [2], [3]. This has created the well-known tussle between users, who strive for fast downloads, and Internet Service Providers (ISPs), whose flat-rate pricing business model is challenged by the enormous p2p traffic volumes. As a result, many ISPs started deploying middle-boxes (e.g., [2], [4]) capable of rate-limiting or blocking p2p protocols [5], [6], [7], [8]. This has in turn created an arms race with p2p developers who started designing methods to evade such restrictions (e.g., [9], [10]).

This problem has drawn a significant attention by the networking research community (e.g., [11], [12], [13], [14], [15], [9], [16], [3]), with several proposals claiming that the tussle between p2p applications and ISPs is unnecessary, and that a “win–win” solution can be achieved via ISP-friendly p2p design. In particular, some proposals show that proximity-based peer selection (e.g., [14]), can help reduce inter-AS traffic volumes. Other proposals (e.g., [15]) argue that cooperation among p2p applications and ISPs is the way to resolve the tussle. Unfortunately, recent research results, i.e., [3], demonstrate that such a win–win outcome is unlikely.

The problem with proximity-based approaches, which we confirm in our work here as well, is that not enough local peers exist in a given region, even for globally popular files. As a result, proximity-only approaches might have limited effects on reducing p2p traffic. The problem with approaches advocating collaboration between p2p applications and ISPs (e.g., [15]) is that ISPs can have divergent needs, and often no incentives to deploy such approaches. We detail more on both approaches in the related work section.

Our key contribution lies in demonstrating that a comprehensive and feasible win–win solution to this emerging problem does exist. In particular, we propose a new service model for p2p file sharing and demonstrate that when this model is applied to BitTorrent, it significantly decreases inter-AS traffic and achieves dramatic p2p performance improvements in the common case. Specifically, we demonstrate that our data delivery approach reduces the inter-AS traffic by 45–75% in the common case. Moreover, it reduces the inter-AS traffic generated by popular torrents by up to an order of magnitude and improves their download times by 60%. Our main goal is to reduce inter-AS traffic volumes without sacrificing peer performance measured by download time.

The common wisdom is that current peer behavior is characterized by limited altruism (willingness to share a file once peers downloaded it), which fundamentally limits the system performance (e.g., [17], [18], [3], [19]). We argue that this is not the case. In particular, existing measurements, including our own, show that the majority of peers stay connected and seed a previously downloaded file several hours after completing a download [3], [19]. We argue that this current level of peer altruism is enormous. Indeed, if peers on average spend two to three times longer as seeders than as leechers (downloaders), this implies that there should be enough resources to effectively serve all peers without any contention. This is true, in the common case, even when we account for asymmetric properties of access networks that typically constrain data transfer. Thus, we argue that the problems in p2p networks, including the inter-AS one, do not arise due to lack of resources, but rather due to the fact that these resources are unmanaged.

Motivated by the effectiveness in which ‘classical’ content distribution networks (CDNs), such as Akamai [20], bring content closer to clients, we propose a system-wide change that can enable effective content replication in swarming p2p systems such as BitTorrent. In particular, our service model enables peers that become seeders to serve the content popular in their own geographic region. This is similar to the way CDNs, such as Akamai, serve content popular in given geographic areas and redirect clients to close-by servers. Thus, instead of seeding a file to a distant peer (and likely getting throttled by intermediate ISPs [5]), the key idea is to reuse precious resources locally.

It is essential to understand that our approach requires no novel incentives for users to change or improve their current sharing behavior. Indeed, users have no incentives to share a file once they download it, yet they do. Our approach simply exploits this existing level of user altruism and further advances the p2p file sharing idea to the next level. Also, one should understand that our approach requires no changes in the way users use p2p file sharing applications. They can join any swarm they like, download any file they want, and disconnect whenever they want. At the same time, our approach does not change the amount of time that users upload, just which data is uploaded to which users. Finally, because our service model enables content to be uploaded to end-user machines, even if they did not originally request it, one concern might be that this can expose users to prosecutable content (for example child pornography). This is avoided in our system via filtering and scheduling policies, as we explain in detail below.