Toward Representative Internet Measurements

Efforts to characterize and understand the structure and behavior of the Internet have a long history in the network research community. Though our understanding of the network has evolved, our knowledge of its configuration, workload, and failure modes still is far from complete. Characterizing the operation of the current Internet infrastructure and its usage patterns is essential to improve the understanding of the Internet and shape its future. Despite the numerous studies of Internet properties, there has been little work to understand how to measure the network in terms of the number and distribution of measurement sites. Research studies that aim to collect and analyze Internet data from end-hosts are typically limited to a relatively small number of probes, located primarily at universities or research institutions. Naturally, the authors of such studies express caveats to their results, and offer some justification of the limited perspective. Some examples of this reasoning in studies of Internet path properties include statements such as “routes between the 37 hosts are plausibly representative” [10], “we believe that measuring the paths between our sources and a wide variety of different ISPs would provide a representative view” [2], and “the testbed topology contains paths that traverse most of the ‘large’ AS’s in the Internet” [8]. Some recent work delves into this issue in the context of discovering the Internet router-level topology [5], concluding that having more than 1 or 2 vantage points does not improve discovery of nodes and links. The authors also point out, however, that it is impossible to claim that the measurement sites used in the study are representative, and that their conclusions are unavoidably sensitive to the choice of measurement sites. While deploying and maintaining a large-scale measurement infrastructure is extremely challenging, it is clear that further guidance as to the representativeness of results obtained using existing measurement testbeds is a crucial missing element of current network measurement research. In this paper we argue that for measurements of traffic, paths, workloads, etc., to be representative, data should be collected from locations actually used by a large portion of Internet users (with an obvious trade-off between “large” and the degree of representativeness). For the common client-server communication pattern, this means that measurement infrastructure should be placed at, or near, places (i.e., stub networks, ingress points, egress points) that are sources or destinations for most of the traffic flowing in the Internet. For popular peer-to-peer applications, measurement probes should be placed at concentration points (i.e., close to busy dial-up access servers or broadband central offices and cable head-ends). Given enough resources, a more complete approach would be to place a measurement probe at every access point in the Internet. This would allow, for example, studies of general properties of the entire network, such as a survey of delays along all paths regardless of their usage. Our view is that though such studies might be valuable, it is more important and interesting, not to mention more practical, to characterize the network based on common usage.

表征和了解互联网的结构和行为在网络研究社区中具有悠久的历史，尽管我们对网络的理解已经发展，但我们对其配置，工作量和故障模式的了解仍然远非完整。属性，几乎没有工作来衡量网络的数量和分布，旨在收集和分析最终主持人的互联网数据，这是相对较少的问题，主要是在大学或研究机构中，这些研究的作者自然而然地表现出了一些效果。互联网路径属性的推理包括诸如“ 37个主机之间的途径是合理的代表性” [10]，“我们相信，测量​​我们的来源和各种不同的ISP之间的路径[2] [2]，以及“互联网大部分互联网”的途径。路由器级别的拓扑[5]得出的结论是，具有1或2个有利位置的结论不会改善节点和链接的发现，但作者也无法指出，研究中使用的测量站点是代表性的，并且他们的结论是对测量的不可避免的，并且在衡量范围内进行了不可避免的构建。关于使用现有测量床获得的结果的进一步指导是当前网络测量研究的关键要素，在本文中，我们认为，对于交通，路径，工作负载等的测量值，应是代表性的，应从互联网用户的大量互联网使用者（与“大型”级别的cirtens servers servers counters semantions contermance contermance contermants contermants countral sembless countral semance of Smointry of Spointers countral and countens courtens semants counters section countens and。模式，这意味着应将基础架构放置在或附近的位置（即存根网络，入口点，出口点），这是大多数流量流量的源或目的地，用于流行的点对点应用程序。方法是在互联网中的每个访问点上放置一个测量探测器，例如，对整个网络的一般属性进行研究，例如对所有路径的调查，无论我们的使用情况如何