Efficient Discovery of Network Topology and Routing Policy in the Internet
Total Page:16
File Type:pdf, Size:1020Kb
Efficient discovery of network topology and routing policy in the Internet Neil Timothy Spring A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy University of Washington 2004 Program Authorized to Offer Degree: Computer Science and Engineering University of Washington Abstract Efficient discovery of network topology and routing policy in the Internet by Neil Timothy Spring Co-Chairs of Supervisory Committee: Associate Professor David J. Wetherall Department of Computer Science and Engineering Professor Thomas E. Anderson Department of Computer Science and Engineering Little is known about the structure and configuration of the ISP networks that constitute the Internet. This shortage of information is a consequence of a fundamental principle of the Internet architecture: that constituent ISPs are administered independently. ISP networks connect by a narrow interface that supports the delivery of data but hides the implementation and behavior of a network from its neighbors. One consequence of this isolation between networks is that, although operators have inside information for their own networks, it is difficult to recognize and fix problems that span administrative boundaries. In this dissertation, I focus on what can be discovered by an outsider: a user or researcher without access to privileged information. I show that the network topologies and routing policies of ISP networks can be discovered despite the narrowness of the interface between them. To do this, I develop and evaluate techniques to measure structured, router-level ISP network topologies and infer intra-domain and peering routing policies. To make these techniques efficient, I use a philosophy of choosing to collect only measurements likely to yield new information. This enables the techniques to run on a large network measurement platform, composed of hundreds of public traceroute servers, to produce an accurate result. I applied and evaluated my techniques to map ten diverse ISP networks and characterize the routing policies of 65 ISP networks. The results are a set of ISP topologies that are several times more complete than previous maps and the first quantitative study of peering routing policy. This data highlights the diversity of ISP networks and can be used by others to better understand network operation and protocol design choices. TABLE OF CONTENTS List of Figures iii List of Tables v Chapter 1: Introduction 1 1.1 Benefits of Measured Topologies and Routing Policies . 2 1.2 Challenges and Goals . 3 1.3 Thesis and Contributions . 5 1.4 Organization . 6 Chapter 2: Background 7 2.1 Internet Topology Concepts . 7 2.2 Internet Routing Policy Concepts . 11 Chapter 3: Related Work 15 3.1 Traceroute and Active Measurement . 15 3.2 BGP Routing Information . 18 3.3 DNS Naming Information . 20 Chapter 4: ISP Topology Mapping 22 4.1 Choosing Traceroutes for Efficient ISP Topology Mapping . 22 4.2 Resolving IP Aliases to Routers . 28 4.3 Decoding DNS Names to Find Structure . 30 4.4 Limitations . 32 4.5 Summary . 33 Chapter 5: ISP Topology Mapping Evaluation 34 5.1 The ISPs . 34 5.2 Efficiency of Traceroute Collection . 35 5.3 Agreement Between Alias Resolution Methods . 42 5.4 Efficiency of Alias Resolution . 46 5.5 Completeness and Consistency of Inferred Geographic Locations . 49 5.6 Accuracy of the Resulting Maps . 53 5.7 Sampling Bias . 56 5.8 Summary . 62 i Chapter 6: Network Routing Policy Inference 63 6.1 Intra-domain Routing Policy Inference . 63 6.2 Intra-domain Routing Model Accuracy . 69 6.3 Measurements for Studying Peering Routing Policy . 73 6.4 Peering Routing Policy Inference . 78 6.5 Peering Routing Model Predictiveness . 82 6.6 Summary . 84 Chapter 7: ISP Maps and Analyses 86 7.1 Measured Router-level ISP Maps . 86 7.2 Analysis of the Measured Router-level ISP Topologies . 89 7.3 Analysis of Routing Policies . 96 7.4 Summary . 103 Chapter 8: Conclusions and Future Work 104 8.1 Thesis and Contributions . 104 8.2 Future Work . 105 Bibliography 109 ii LIST OF FIGURES Figure Number Page 2.1 An illustration of an inter-domain topology. 8 2.2 An illustration of a POP-level topology. 8 2.3 An illustration of a small office network topology. 10 2.4 A sample BGP routing table snippet. 12 2.5 Early- and late-exit peering routing policies. 13 3.1 A sample of traceroute output. 16 3.2 Internet map visualizations from different projects. 17 3.3 Traceroute in opposite directions. 18 4.1 Path reductions find traceroutes likely to traverse the same path. 25 4.2 The Rocketfuel topology-mapping engine data-flow. 27 4.3 Alias resolution: recognizing interfaces that belong to the same router. 28 4.4 Alias resolution using IP identifiers. 29 4.5 A sample naming convention rule from undns. 32 5.1 Ingress reduction: vantage points share ingresses. 38 5.2 Egress reduction: dependent prefixes share egresses. 39 5.3 Next-hop AS reduction: prefixes share next-hop ASes. 42 5.4 Completeness of source-address-based alias resolution from many points. 46 5.5 Distribution of return TTLs from discovered addresses. 47 5.6 Distribution of the difference between return TTLs. 48 5.7 Distribution of the Euclidean distance between return TTL coordinates. 49 5.8 Distribution of aliases found for increasingly different DNS names. 50 5.9 Geographic locations known to the undns library. 51 5.10 Comparison of BGP adjacencies with Route Views. 55 5.11 Comparison with Skitter by ISP. 56 5.12 Router out-degree by ISP, sampling bias analysis of near and far sets. 60 6.1 Example weighted intra-domain topology. 65 6.2 Eliminating redundant constraints. 66 6.3 Percentage of observed paths that were least cost. 70 6.4 Percentage of dominant paths that were least cost. 71 6.5 Percentage of node-pairs fully and partially modeled. 72 6.6 Predictive power of routing model for Exodus. 72 6.7 The link weights of stub connections are undefined. 78 6.8 Inferred weights: comparison to latency model. 79 6.9 Inferred weights: predictive from a subset of paths. 79 6.10 Different peering policies affect how peering points are found. 83 iii 7.1 Measured backbone topologies of US ISPs. 87 7.2 A sample POP topology from Sprint. 88 7.3 Router out-degree distribution, aggregated. 90 7.4 Router out-degree distribution, by ISP. 91 7.5 Backbone router out-degree distribution, aggregated. 92 7.6 Distribution of POP sizes and routers in POPs. 93 7.7 Backbone routers in a POP relative to its size. 94 7.8 POP out-degree vs. backbone routers in the POP. 94 7.9 POP out-degree distribution. 95 7.10 Distribution of router-level adjacencies for each AS adjacency. 96 7.11 Distribution of external adjacencies per POP. 97 7.12 Path inflation due to intra-domain topology. 98 7.13 Path inflation due to intra-domain routing policy. 99 7.14 The prevalence of early-exit routing. 100 7.15 Fraction of non-early-exit paths routed closer to the destination. 101 iv LIST OF TABLES Table Number Page 5.1 The number of routers, links and POPs for all ten ISPs studied. 35 5.2 Directed probing effectiveness. ..