A probabilistic approach to building large-scale federated systems

Rising Internet connectivity and emerging web service standards are enabling a new federated computing model, where computing systems will be comprised of multiple components distributed across multiple collaborative organizations. While federated services can revolutionize collaboration and commerce across the Internet, the realization of this vision faces a number of challenges arising from its fundamental cross-organizational nature. In addition, problems like faulty hardware and software, operator error, dynamic reallocation of resources, load spikes, and malicious users cause these systems to be highly volatile making federated computing even more challenging. Traditional techniques for building distributed systems, have generally provided resource management and communication by relying on structured solutions like: (a) imposing an overlay structure over the system (i.e. multicast tree or distributed hash table), (b) depending on centralized services or (c) relying on distributed consistency protocols. Previous work [42,46,64,102], has shown that these techniques become expensive and sometime unfeasible in environments where membership changes rapidly and nodes are heterogeneous and unpredictable. In this dissertation, we explore a different approach for building large scale distributed systems. Our thesis is to create distributed algorithms that allow members to operate autonomously so that their progress is not conditioned by other nodes. Despite their independence, as a whole members should be able to make constant progress toward achieving stable global goals. In order to ensure autonomy, global progress and stability, we build randomized algorithms that depend only on loosely synchronized global data. In this dissertation we explore an infrastructure called PlanetP that we have simulated and partially prototyped to validate our thesis. PlanetP embodies several of our ideas by using probabilistic algorithms to provide, group communication, membership management, content based search and ranking, autonomous service deployment and management, and autonomous replication to provide predictable data availability. Our work is novel in that we target highly dynamic environments where nodes join and leave constantly in an uncontrolled manner.