Design and evaluation of distributed algorithms for placement of network services.

摘要

Network services play an important role in the Internet today. They serve as data caches for websites, servers for multiplayer games and relay nodes for Voice over IP (VoIP) conversations. While much research has focused on the design of such services, little attention has been focused on their actual placement. This placement can impact the quality of the service, especially if low latency is a requirement. These services can be located on nodes in the network itself, making these nodes supernodes. Typically supernodes are selected in either a proprietary or ad hoc fashion, where a study of this placement is either unavailable or unnecessary. Previous research dealt with the only pieces of the problem, such as finding the location of caches for a static topology, or selecting better routes for relays in VoIP. However, a comprehensive solution is needed for dynamic applications such as multiplayer games or P2P VoIP services. These applications adapt quickly and need solutions based on the immediate demands of the network.;In this thesis we develop distributed algorithms to assign nodes the role of a supernode. This research first builds off of prior work by modifying an existing assignment algorithm and implementing it in a distributed system called Supernode Placement in Overlay Topologies (SPOT). New algorithms are developed to assign nodes the supernode role. These algorithms are then evaluated in SPOT to demonstrate improved SN assignment and scalability. Through a series of simulation, emulation, and experimentation insight is gained into the critical issues associated with allocating resources to perform the role of supernodes. Our contributions include distributed algorithms to assign nodes as supernodes, an open source fully functional distributed supernode allocation system, an evaluation of the system in diverse networking environments, and a simulator called SPOTsim which demonstrates the scalability of the system to thousands of nodes. An example of an application deploying such a system is also presented along with the empirical results.