Overcast is an application-level multicasting system that can be incrementally deployed using today's Internet infrastructure. These properties stem from Overcast's implementation as an overlay network. An overlay network consists of a collection of nodes placed at strategic locations in an existing network fabric. These nodes implement a network abstraction on top of the network provided by the underlying substrate network.

The Internet routing fabric is partitioned into several domains. Each domain represents a region of the fabric administered by a single commercial entity. Over the past two years, the routing fabric has experienced significant growth. From more than a year's worth of inter-domain routing traces, we analyze the Internet inter-domain topology, its route stability behavior, and the effect of growth on these characteristics. Our analysis reveals several interesting results. Despite growth, the degree distribution and the diameter of the inter-domain topology have remained relatively unchanged. Furthermore, there exists a four-level hierarchy of Internet domains classified by degree. However, connectivity between domains is significantly nonhierarchical. Despite increased connectivity at higher levels in the topology, the distributionof paths to prefixes from the backbone remained relatively unchanged. There is evidence that both route availability and the mean reachability duration have de...

We present a natural greedy algorithm for the metric uncapacitated facility location problem and use the method of dual fitting to analyze its approximation ratio, which turns out to be 1.861. The running time of our algorithm is O(m log m), where m is the total number of edges in the underlying complete bipartite graph between cities and facilities. We use our algorithm to improve recent results for some variants of the problem, such as the fault tolerant and outlier versions. In addition, we introduce a new variant which can be seen as a special case of the concave cost version of this problem.

Distributed Hash Tables have been proposed as flat, nonhierarchical structures, in contrast to most scalable distributed systems of the past. We show how to construct hierarchical DHTs while retaining the homogeneity of load and functionality offered by flat designs. Our generic construction, Canon, offers the same routing state v/s routing hops trade-off provided by standard DHT designs. The advantages of Canon include (but are not limited to) (a) fault isolation, (b) efficient caching and effective bandwidth usage for multicast, (c) adaptation to the underlying physical network, (d) hierarchical storage of content, and (e) hierarchical access control. Canon can be applied to many different proposed DHTs to construct their Canonical versions. We show how four different DHTs—Chord, Symphony, CAN and Kademlia—can be converted into their Canonical versions that we call Crescendo, Cacophony,

Large-scale distributed systems require new middleware paradigms that do not suffer from the limitations of traditional request/reply middleware. These limitations include tight coupling between components, a lack of information filtering capabilities, and support for one-to-one communication semantics only. We argue that event-based middleware is a scalable and power-ful new type of middleware for building large-scale distributed systems. However, it is important that an event-based middleware platform includes all the standard functionality that an appli-cation programmer expects from middleware. In this thesis we describe the design and implementation of Hermes, a distributed, event-based middleware platform. The power and flexibility of Hermes is illustrated throughout for two application domains: Internet-wide news distribution and a sensor-rich, active building. Hermes follows a type- and attribute-based publish/subscribe model that places particular emphasis on programming language integration by supporting type-checking of event data and event type inheritance. To handle dynamic, large-scale environments, Hermes uses peer-to-peer techniques for autonomic management of its overlay network of event brokers and for scalable

End-system multicast provides a low-cost solution to scalably broadcast information to groups of users. However, lastmile bandwidth limitations constrain tree fanouts leading to high end-to-end delivery delays. These delays can be reduced if the network provides forwarding proxies with high fanout capabilities at an additional cost. We use simple graph theoretic network models to explore the problem of building hybrid proxy/end-system application layer multicast trees that meet fixed end-to-end delay bounds. Our goal is to meet a fixed delay bound while minimizing costs associated with the utilization of proxies. We provide an algorithm and formally prove its optimality in a fully-connected overlay network with uniform-length edges. We then adapt this algorithm into a heuristic and evaluate the heuristic for simulated transit-stub networks with variable-delay edges. We compare our heuristic in a proxy-free environment to previously developed heuristics and show that our heuristic typically yields further reductions in the maximum session end-to-end delay.

This thesis will describe an approach to supporting overlay networks in the Internet 's network layer. Two primitives, Packet Reflection and Path Painting, will be described and evaluated.

Abstract Widespread replication of information can ameliorate the problem of server overloading but raises the allied question of server selection. Clients may be assigned to a replica in a static manner or they may choose among replicas based on client-initiated measurements. The latter technique, called dynamic server selection (DSS), can provide significantly improved response time to users when compared with static server assignment policies (for example, based on network distance in hops). In the first part of this paper we demonstrate the idea of DSS using experiments performed in the Internet. We compare a range of policies for DSS and show that obtaining additional information about servers and paths in the Internet before choosing a server improves response time significantly. The best policy we examine adopts a strategy of never adding more than one percent additional traffic to the network, and is still able to provide nearly all the benefits of the most expensive policies. While these results suggest that DSS is beneficial from the network user's standpoint, the system-wide effects of DSS schemes should also be closely examined. In the second part of this paper we use large-scale simulation to study the system-wide network impact of dynamic server selection. We use a simulated network of over 100 hosts that allows local-area effects to be distinguished from wide-area effects within traffic patterns. In this environment we compare DSS with static server selection schemes and confirm that client benefits remain even when many use DSS simultaneously. Importantly, we also show that DSS confers system-wide benefits from the network standpoint, as compared to static server selection. First, overall data traffic volume in the network is reduced, since DSS tends to diminish network congestion. Second, traffic distribution improves--\Lambda This work was done while the author was at Boston University.

Of the many distributed applications designed for the Internet, the successful ones are those that have paid careful attention to scale and robustness. These applications share several design principles. In this paper, we illustrate the application of these principles to common network monitoring tasks. Specifically, we describe and evaluate 1) a robust distributed topology discovery mechanism and 2) a mechanism for scalable fault isolation in multicast distribution trees. Our mechanisms reveal a different design methodology for network monitoring one that carefully trades off monitoring fidelity (where necessary) for more graceful degradation in the presence of different kinds of network dynamics.

In a Web site, each page v has a probability...