Resumen de On the scalability of LISP and advanced overlaid services
Florin Tudorel Coras
In just four decades the Internet has gone from a lab experiment to a worldwide, business critical infrastructure that caters to the communication needs of almost a half of the Earth's population. With these figures on its side, arguing against the Internet's scalability would seem rather unwise. However, the Internet's organic growth is far from finished and, as billions of new devices are expected to be joined in the not so distant future, scalability, or lack thereof, is commonly believed to be the Internet's biggest problem. While consensus on the exact form of the solution is yet to be found, the need for a semantic decoupling of a node's location and identity, often called a location/identity separation, is generally accepted as a promising way forward. Typically, this requires the introduction of new network elements that provide the binding of the two names-paces and caches that avoid hampering router packet forwarding speeds. But due to this increased complexity the solution's scalability is itself questioned. This dissertation evaluates the suitability of using the Locator/ID Separation Protocol (LISP), one of the most successful proposals to follow the location/identity separation guideline, as a solution to the Internet's scalability problem. However, because the deployment of any new architecture depends not only on solving the incumbent's technical problems but also on the added value that it brings, our approach follows two lines. In the first part of the thesis, we develop the analytical tools to evaluate LISP's control plane scalability while in the second we show that the required control/data plane separation provides important benefits that could drive LISP's adoption. As a first step to evaluating LISP's scalability, we propose a methodology for an analytical analysis of cache performance that relies on the working-set theory to estimate traffic locality of reference. One of our main contribution is that we identify the conditions network traffic must comply with for the theory to be applicable and then use the result to develop a model that predicts average cache miss rates. Furthermore, we study the model's suitability for long term cache provisioning and assess the cache's vulnerability in front of malicious users through an extension that accounts for cache polluting traffic. As a last step, we investigate the main sources of locality and their impact on the asymptotic scalability of the LISP cache. An important finding here is that destination popularity distribution can accurately describe cache performance, independent of the much harder to model short term correlations. Under a small set of assumptions, this result finally enables us to characterize asymptotic scalability with respect to the amount of prefixes (Internet growth) and users (growth of the LISP site). We validate the models and discuss the accuracy of our assumptions using several one-day-long packet traces collected at the egress points of a campus and an academic network. To show the added benefits that could drive LISP's adoption, in the second part of the thesis we investigate the possibilities of performing inter-domain multicast and improving intra-domain routing. Although the idea of using overlaid services to improve underlay performance is not new, this dissertation argues that LISP offers the right tools to reliably and easily implement such services due to its reliance on network instead of application layer support. In particular, we present and extensively evaluate Lcast, a network-layer single-source multicast framework designed to merge the robustness and efficiency of IP multicast with the configurability and low deployment cost of application-layer overlays. Additionally, we describe and evaluate LISP-MPS, an architecture capable of exploiting LISP to minimize intra-domain routing tables and ensure, among other, support for multi protocol switching and virtual networks.
