Peer-to-peer (P2P) locality has recently raised a lot of interest in the community. Indeed, whereas P2P content distribution enables financial savings for the content providers, it dramatically increases the traffic on inter-ISP links. To solve this issue, the idea to keep a fraction of the P2P traffic local to each ISP was introduced a few years ago. Since then, P2P solutions exploiting locality have been introduced. However, several fundamental issues on locality still need to be explored. In particular, how far can we push locality, and what is, at the scale of the Internet, the reduction of traffic that can be achieved with locality? In this paper, we perform extensive experiments on a controlled environment with up to 10,000 BitTorrent clients to evaluate the impact of high locality on inter-ISP links traffic and peers download completion time. We introduce two simple mechanisms that make high locality possible in challenging scenarios and we show that we save up to several orders of magnitude inter-ISP traffic compared to traditional locality without adversely impacting peers download completion time. In addition, we crawled 214,443 torrents representing 6,113,224 unique peers spread among 9,605 ASes. We show that whereas the torrents we crawled generated 11.6 petabytes of inter-ISP traffic, our locality policy implemented for all torrents could have reduced the global inter-ISP traffic by up to 40%.
Content distribution is today at the core of the services provided by the Internet. However, distributing content to a large audience is costly with a classical client-server or CDN solution. This is the reason why content providers start to move to P2P content distribution that enables to significantly reduce their cost without penalizing the experience of users. One striking example is Murder, a BitTorrent extension to update the Twitter infrastructure.
However, whereas current P2P content distribution solutions like BitTorrent are very efficient, they generate a huge amount of traffic on inter-ISP links [1]. Indeed, in BitTorrent, each peer that downloads a given content is connected to a small subset of peers picked at random among all the peers that download that content. In fact, even though peers in the same ISP are downloading the same content they are not necessarily connected to each other. As a consequence, peers unnecessarily download most of the content from peers located outside of their ISP.
Therefore, even if current P2P content replication solutions significantly reduce content provider costs, they cannot be promoted as a global solution for content replication as they induce huge costs for ISPs. In particular, the current trend for ISPs is to block P2P traffic [2].
One solution to this problem is to use P2P locality. The goal Email addresses: stevens.le_blond@inria.fr (Stevens Le Blond), arnaud.legout@inria.fr (Arnaud Legout), walid.dabbous@inria.fr (Walid Dabbous) of P2P locality is to constrain P2P traffic within ISPs’ boundaries in order to minimize the amount of inter-ISP traffic.
The seminal work of Karagiannis et al. [1] is the first one to suggest the use of locality in a P2P system in order to reduce the load on inter-ISP links. They show on real traces the potential for locality (in particular spatial and temporal correlation in the requests for contents) and, based on simulation on a BitTorrent tracker log, they evaluate the benefit of several architectures and in particular a P2P architecture exploiting locality. More recently, Xie et al. [3] proposed P4P, an architecture to enable cooperation between P2P applications and ISPs. They show by performing large field tests that P4P enables reduction of external traffic for a monitored ISP and enables a reduction on the peers download completion time. Choffnes et al. [4] proposed Ono, a BitTorrent extension that leverages on a CDN infrastructure to localize peers in order to group peers that are close to each other. They show the benefit of Ono in terms of peers download completion time and that Ono can reduce the number of IP hops and AS hops among peers.
With those works, there is no doubt that P2P locality has some benefits and that there are several ways to implement it. However, two fundamental questions are left unanswered by those previous works.
How far can we push locality? In all proposed solutions the number of inter-ISP connections is kept high enough to guarantee a good robustness to partitions, i.e., a lack of connectivity among set of peers resulting in a poor download completion time. However, this robustness is at the expense of a larger inter-ISP traffic. How far can we push locality without impact-ing the robustness to partition of the P2P protocol?
What is, at the scale of the Internet, the reduction of traffic that can be achieved with locality? It might be argued that P2P locality will bring little benefits at the scale of the Internet. Indeed, in case most ISPs have just a few peers, there will be little inter-ISP traffic reduction by keeping the traffic local to those ISPs. Therefore, the question is, what is the distribution of peers per ISP in the Internet, and what would be the inter-ISP bandwidth savings achieved with a locality policy. Previous works looking at inter-ISP bandwidth savings either consider indirect measurements (like the distribution of the number of AS between any two peers with a direct connection [4]), partial measurements (like the monitoring of a specific ISP), or simulations (like comparing various content distribution scenarios based on the location of peers obtained from a tracker log). For instance, Xie et al. [3] reported results on inter-ISP savings with P4P for a single ISP.
The answers to those questions will be fundamental when P2P content replication will be used by content providers for large scale distribution. In that case, it is likely that ISPs will need to know the amount of inter-ISP traffic they can save with locality, and that they will request content providers to minimize this traffic due to P2P applications accordingly. At the same time, the content providers will need a clear understanding of the impact of this reduction of traffic on their customers.
Our contribution in this paper is to answer those questions by running extensive large scale BitTorrent experiments (with up to 10 000 real BitTorrent clients) in a controlled environment, and by using real
This content is AI-processed based on open access ArXiv data.
