Having so many independent generation models and topology generators is disadvantageous in many respects. A researcher in need of synthetic topologies to investigate the correctness and performance of protocols and algorithms is forced to learn the nuances of many of these models/generators. Consequently she may be forced to use the most popular one, the one supported in the simulation environment used, or the easiest one. As we mentioned before, different generators produce topologies that are aimed to be used in different contexts and with different goals. For example, it does not make too much sense to use BRITE 1.0, Inet or PLRG to generate 20-node topologies for simulations. Similarly, if the AS-level connectivity properties (e.g. degree-related characteristics) are an important consideration, then using GT-ITM may not be appropriate. The result is that the researcher may end up having to use more than one generator or use one that does not offer important properties in the generated topologies (hierarchy in some cases, power-laws in others). Analogously, for a researcher investigating the challenges of topology generation and looking for better and more powerful generation models, having so many generators available makes comparative analyses of different models significantly more difficult. For example, in order to perform a thorough comparative study one may have to learn to use GT-ITM, BRITE 1.0, Inet, PLRG and others, as well as to understand the different output formats, write different filtering routines for different output files, etc. Furthermore, if a new generation model is envisioned, a researcher has two options. Either a new generator is developed or an existing one is extended. Clearly, developing a new generator is cumbersome and available generators are not designed to be easily extended or even modified.
These difficulties are in addition to the inherently hard problems encountered when developing models that accurately capture fundamental properties of the Internet topology. Such a model is usually developed based on actual topological information that is not completely accurate. Such lack of accuracy is mainly due to the fact that mapping the Internet topology is a very challenging task [11,23]. At the Autonomous System (AS) level, available information is richer because it can be obtained or inferred from BGP tables [17,10]. In contrast, accurate router-level topological information is hard to obtain and until now inferring router-level connectivity has been done by using traceroute or traceroute-like probing mechanisms [11,6]. Identifying the actual fundamental properties of topologies at the router-level is still an open research question . Most Internet topology studies have approached topology modeling relying only on physical connectivity. However, routing in the Internet is determined by a policy-based routing protocol (BGP) and consequently physical connectivity does not always implies reachability. Customer-provider and peering relationships play a deciding role in determining whether or not traffic can flow between connected nodes. As an example, consider a customer AS that is connected to two provider ASs, AS-1 and AS-2. In general, a customer AS does not provide transit between its providers. So, even though there is a path from provider AS-1 to provider AS-2, they may not actually exchange traffic via the customer AS. Hence the connectivity of a topology alone does not completely characterize the structural properties of the corresponding routing topology . Even if we knew the actual relationships between ASs, such relationships are continuously changing. Therefore, in order to generate accurate representative topologies the invariants of such relationships across time and size must be discovered.
In short, research in topology generation is in its infancy. New models will be developed as research will expose new and more powerful mechanisms to accurately characterize the topology of the Internet. topologies. Our challenges can be concisely put into two issues: