Section: Overall Objectives

Resource management in peer-to-peer overlays

Managing resources on a large scale, be them computing resources, data, events, bandwidth, requires a fully decentralized solution. Our research in this area focuses on building the relevant overlay networks to provide core functionalities of resource management and discovery. This includes broadcast, anycast, search, notification. Overlay networks organize peers in a logical network on top of an existing networking infrastructure. The system automatically and dynamically adapts to frequent peer arrivals and departures. Two main classes have been designed. (1) Structured overlay networks, such as distributed hash tables (DHT), rely on a logical structure and map object keys to overlay nodes. While structured networks were initially more popular among scientists, their exact-match interface limits their flexibility and use for various applications, notably for non-exact information retrieval. (2) Unstructured overlay networks connect peers randomly (or pseudo-randomly). This class of networks is dominated by broadcast-based searching techniques, where the goal has become to enforce restrictions on broadcasting so that efficiency can be guaranteed.

In this area, our approach is original for three reasons. First, we are focusing on creating overlays taking into account application characteristics. This translates into either connecting applications objects themselves as peers (which obviously are eventually hosted on a physical computing entity), or influencing the overlay links so that the structure of the application itself can be leveraged for a better performance. We strongly believe that it is not possible to rely on a generic framework applicable to all potential large-scale platforms (as the Internet, grids, or wireless autonomic networks), because of the heterogeneity and limitation of resources among participating entities.

Second, we strongly believe in weakly-structured networks and most of our projects rely on epidemic-based unstructured overlays. Epidemic communication models have recently started to be explored as a general paradigm to build and maintain unstructured overlay networks. The basic principle of such epidemic protocols is that periodically, each peer exchanges information with some other peers selected from a local list of neighbors. Such protocols have shown to be extremely resilient to network dynamics [100] .

Finally, we are convinced that we can greatly benefit from the experience gathered from both existing systems and theoretical models. We spend a significant amount of energy to find, gather and analyze workloads of real systems as well as to develop our own platform in the context of our peer-to-peer collaborative backup platform.