BitTorrent Experiments on Testbeds: A Study of the Impact of Network Latencies

arXiv:1003.5746v1 [cs.NI] 30 Mar 2010 BitTorrent Experiments on Testbeds: A Study of the Impact of Network Latencies Ashwin Rao, Arnaud Legout, and Walid Dabbous INRIA, France. {ashwin.rao, arnaud.legout, walid.dabbous}@inria.fr Abstract—In this paper, we study the impact of network latency on the time required to download a file distributed using BitTorrent. This study is essential to understand if testbeds can be used for experimental evaluation of BitTorrent. We observe that the network latency has a marginal impact on the time required to download a file; hence, BitTorrent experiments can performed on testbeds. I. INTRODUCTION Testbeds such as PlanetLab and Grid5000 are widely used to study the performance of communication protocols and networking applications. One commonly used practice while performing experiments on such testbeds is to run multiple in- stances of the application being studied on the same machine. However, one primary shortcoming of this approach is the absence of any network latency between the instances of the application running on the same machine. Further, in experi- ments involving more than one machine, the latency between the machines present in the same local area network (LAN) is negligible. In this paper we study the impact of network latency on the outcome of experiments that are performed on testbeds to evaluate the performance of BitTorrent. The BitTorrent Protocol internally uses the Transmission Control Protocol (TCP) while distributing the content [1]. The steady-state throughput of TCP is function of the round- trip time (RTT) [2]. Further, the slow start and congestion avoidance phase of TCP introduce a ramp up period which is required to attain a throughput equal to the minimum of the network throughput and the rate at which the application is sending data. This ramp up period is a function of the RTT and the rate at which the data is being uploaded. BitTorrent allows the users to limit the rate at which data is uploaded; as the time duration of an upload by a peer is in the order of seconds, we believe that the time required to transfer pieces of a file is not affected by such variations in the TCP throughput. Our experiments show that the RTT (and hence the latency) between the peers in the torrent has a marginal impact (less than 15%) on the time required to download a file. The details of the methodology and the tools used are pre- sented in Section II. We initially assume the latency between any two peers in the torrent to be the same (homogeneous latency); the impact of homogeneous latency on the time required to download a file are presented in Section III. The results without this assumption are presented in Section IV, followed by the conclusions in Section V. II. METHODOLOGY In this paper we use the terminology used by the BitTorrent community. A torrent, also known as a BitTorrent session or a swarm, consists of a set of peers that are interested in having a copy of the given content. A peer in a torrent can be in two states: the leecher state when it is downloading the contents, and the seed state when it has a copy of the content being distributed. A tracker is a server that keeps track of the peers present in the torrent. A. Experiment Scenarios We consider a torrent consisting of one tracker and a finite number of peers; a few of these peers are seeds, while the rest of the peers are leechers. We assume that the peers remain in the torrent until all the leechers have finished downloading the file. The metric used to study the impact of the network latency between the peers is the download completion time, the time required to the download the file distributed using BitTorrent. We use the following network topologies to evaluate the impact of latency on download completion time of a file. 1) Homogeneous Latency. The latency between any two peers in the torrent is the same in this network topology. This topology provides an upper bound on the download completion time when the maximum round trip time between the peers in a torrent is known. Further, this setting was used to give an insight on the threshold of the latency between the peers beyond which the latency affects the download completion time. 2) Heterogeneous Latency. The peers are grouped together to abstract Autonomous Systems (AS). We assume that the latency between any two peers in a given AS is the same and that all ASes are fully meshed. Further, we assume that the inter-AS latency is greater than the intra- AS latency; we also assume symmetric latency in the upload and download links within an AS and between ASes. All the experiments were performed in a private torrent consisting of one tracker, one initial seed (henceforth called as the seed), and 300 leechers; these experiments were carried out on the Grid’5000 experimental testbed [3]. A 50 MB file was distributed in this torrent where the upload rates of the leechers and the seed was varied from 10 kB/s to 100 kB/s. As shown in Figure 1, four

…(Full text truncated)…

This content is AI-processed based on ArXiv data.