INTERNET GLOBAL STRUCTURES
One of the basic tenets of network science is that the structure of a network not only shapes the behavior of its components but also contributes to emergent properties at an aggregate level.
Important features are reliability and robustness, the ability to absorb dysfunctions, and the capacity to survive partial failures. Simulation studies in this research tradition show that both properties are functions of the network structure. In the following paragraphs we will therefore first discuss studies on various network topologies, and then address the issue of network robustness.4.4.1 Network Topologies
Network science has been widely used to study the topology of the physical network infrastructures of the Internet and the economic and social networks formed on them. Using graph theory concepts outlined above, and strongly influenced by the work of Erdos and Reny, network scientists have highlighted the difference between regular and random networks in contrast to scale-free networks that are governed by power laws. The term ‘scale free’ is derived ‘from the simple observation that power-law node degree distributions are free of scale - most nodes have small degree, a few nodes have very high degree, with the result that the average node degree is essentially non-informative’ (Willinger et al., 2010, p. 588).
As we have outlined above, in regular networks all nodes have an equal degree, in random networks it depends on the specific network model. In all networks relative degree frequency can be taken as the probability that a randomly chosen node will have a particular number of links. In a normal distribution the relative degree frequency follows a bell curve in which the average degree lies at the peak of the distribution, but in a powerlaw distribution the curve begins with a peak and decreases exponentially.
Network scientists applied these concepts in empirical studies of real-world big data in nature, society, and technology and discovered that many relations exhibit powerlaw distributions.
Similar structures were found in various social networks, although the findings suggest considerable diversity (Adamic and Adar, 2003; Jackson, 2008; Easley and Kleinberg, 2010; Leskovec et al., 2011). Despite the pervasive presence of power laws their universal application to communication networks and their usefulness remain under debate (Clegg et al., 2010). Simon Knight et al. (2011, p. 10) describe a dataset that includes information on 232 networks. The authors demonstrate a broad variety of network topologies in the ‘Internet topology zoo’. They conclude that there is:[...] evidence for or against many phenomena, for instance: against power-law degree distributions; for hub-and-spoke like behavior; for hierarchy; but the evidence is never completely convincing, reflecting the sheer variety in the networks. If there is any message in this data, it is that there are as many types of networks as there are network designers.
There is a widespread perception that the Internet is a fully connected and widely open virtual space. However, a closer examination of the link structure of the Internet, the World Wide Web, and of other services reveals a more multi-faceted picture. Such a topology can be constructed from various aspects of the Internet, including the link structure among nodes, the flow of data, the control arrangements among autonomous systems (AS), and the management of the Internet. Research dating back to the pioneering work of Broder et al. (2000) reveals that at the core of the Internet is a giant strongly connected component (SCC) within which sites are mutually reachable. In addition, there is a large in-component (sites that are linked to the SCC but cannot be reached from it) and an out-component (sites that can be reached from the SCC but are not connected to it). Furthermore, there are ‘tendrils’ that are linked to the in- and out-components and disconnected components. Work at the Center for Applied Internet Data Analysis (CAIDA) represents these structures in concentric circles outward from the nodes with highest degree, creating jellyfish-like visualizations1 that reveal a similar multi-faceted topology.
4.4.2 Internet Robustness
An important discovery of network science was the effect of alternative network topologies on network robustness in cases of accidental failure or deliberate attack.
Robust networks continue to function even if some nodes or components fail, and experience blackouts only in extreme situations. Robust networks are also highly resilient in that they quickly recover after a serious breakdown. While the distributed nature of the Internet is inherently designed for robustness - in contrast to traditional hierarchical telecommunication networks - the degree of robustness ultimately depends on the topology, that is, the layout of a network. Network research methods can convincingly answer questions regarding which kind of network structures are particularly robust.A conventional wisdom in organizational research is that redundant structures are less prone to disruption. In an overview Mitchell (2009) points to a similar observation, summarizing that a very important property of scale-free networks is their resilience to the deletion of nodes. If a set of random nodes and their links are deleted from a large scale- free network, the network’s basic properties would not change. However, Albert et al. (2000) show that scale-free networks also imply some vulnerabilities. Using formal modeling and computer simulation, they demonstrate that not all redundant systems but only scale-free networks (such as the Internet and social and biological networks) are error tolerant. However, they also find that error tolerance of these networks comes at a price as they are extremely vulnerable to attacks targeting and deliberately removing a few central nodes. Robustness is studied by the simulation of random removal of nodes in different network structures and its effects on connectedness. The latter is measured as change in the diameter (i.e., the shortest path between the two most distant nodes in a network) of a given network as a function of the fraction of the removed nodes. Albert et al. (2000) compare a heterogeneous scale-free network with a homogeneous Erdos-Reny (ER) network in which most nodes have approximately the same number of links.
When nodes are removed, many groups of nodes are cut off from the main cluster. When nodes with the highest degree are eliminated, the diameter of the scale-free network increases rapidly. It doubles its original value if only 5 percent of the nodes are removed. This vulnerability stems from the heterogeneous degree distribution, where connectivity is based on a few highly connected nodes. Their overall finding is that scale- free networks display a surprisingly high degree of tolerance against random failures, a property not shared by the ER network. But error tolerance would come at the price of attack survivability. This finding highlights topological weaknesses of the current communication networks.
Crucitti et al. (2004) followed a similar research track with a more extended methodology using also a centrality concept developed in social network analysis. The effects of errors and attacks on different network topologies have been studied with respect to the efficiency of a graph, that is, the length of communication-link sequences measured as the average geodesic path length. Removal of nodes was simulated with respect to accidental errors but also deliberate attacks on central nodes, where centrality was measured by degree and betweenness centrality. Their findings were similar to those by Albert et al. (2000): homogeneous ER graphs exhibit some tolerance with respect to errors and attacks whereas heterogeneous scale-free networks are robust to errors but vulnerable to attacks. Crucitti et al. (2004) conclude that great effort would be necessary in order to protect many real-world networks such as the Internet from attacks.
4.5