The current Internet includes a large number of distributed services. In order to guarantee the QoS of the communication in these services, a client has to select a close-by server with enough available resources. In order to achieve this objective, in this Thesis, we propose a simple and practical solution for Dynamic and Location Aware Server Discovery based on a Distributed Hash Table (DHT). Specifically, we decide to use a Chord DHT system (although any other DHT scheme can be used). In more detail, the solution works as follows. The servers offering a given service form a Chord-like DHT. In addition, they register their location (topological and/or geographical) information in the DHT. Each client using the service is connected to at least one server from the DHT. Eventually, a given client realizes that it is connected to a server providing a bad QoS, then, it queries the DHT in order to find an appropriate server (i.e. a close-by server with enough available resources). We define 11 design criteria, and compare our solution to the State of the Art based on them. We show that our solution is the most complete one. Furthermore, we validate the performance of our solution in two different scenarios: NAT Traversal Server Discovery and Home Agent Discovery in Mobile IP scenarios. The former serves to validate our solution in a highly dynamic environment whereas the latter demonstrates the appropriateness of our solution in more classical environments where the servers are typically hosts.

The extra overhead suffered from the servers involved in our system comes from their participation in the Chord DHT. Therefore, it is critical to fairly balance the load among all the servers. In our system as well as in other P2P systems (e.g. P2PSIP) the stored objects are small, then routing dominates the cost of publshing and retrieving objects. Therefore, in the second part of this Thesis, we address the issue of fairly balancing the routing load in Chord DHTs. We present an analytical model to evaluate the routing fairness of Chord based on the well accepted Jain’s Fairnes Index. Our model shows that Chord performs poorly. Following this observation, we propose a simple enhancement to the Chord finger selection algorithm with the goal of mitigating this effect. The key advantage of our proposal as compared to previous approaches is that itadds a neglible overhead to the basic Chord algorithm. We validate the goodness of the proposed solution analytically and by large scale simulations.

The conference will be conducted in English