InterestCast: Adaptive Event Dissemination for Interactive Real-Time Applications.
Technische Universität Darmstadt, Darmstadt
[Ph.D. Thesis], (2016)
Thesis 2016-04-18 tuprints.pdf - Accepted Version
Available under Creative Commons Attribution Non-commercial No-derivatives 3.0 de.
Download (4MB) | Preview
|Item Type:||Ph.D. Thesis|
|Title:||InterestCast: Adaptive Event Dissemination for Interactive Real-Time Applications|
Many networked applications use push-based many-to-many communication. Especially real-time applications, such as online games and other virtual reality applications, need to synchronize state between many participants under strict latency requirements. Those applications typically exchange frequent state updates and therefore require an appropriate dissemination mechanism. Centralized and server-based solutions do not minimize latency, since they always need an extra round-trip to the server. In addition, a server infrastructure constitutes a potential performance bottleneck and thus a scalability limitation. Direct communication between event source and destination is often latency-minimal but quickly exceeds the capacities especially of poorly connected participants because each one needs to communicate individually with many others. Our proposed solution, InterestCast, provides a decentralized event dissemination mechanism that uses peer-to-peer event forwarding, allowing powerful participants to help weak participants with the event multiplication and dissemination. To obtain forwarding configurations that best fit the current situation and application needs, InterestCast adapts them dynamically and continuously during runtime. The application’s needs are passed as utility functions, which determine the utility of events with a given latency for a given interest level. Interest levels serve as an abstraction for the importance of events from a specific source, allowing a more fine-grained prioritization than an all-or-nothing subscription model. This is particularly useful if the importance of updates depends on virtual reality distance or another application-specific metric. InterestCast runs an incremental local optimization algorithm that repeatedly evaluates all possible rerouting operations from the point of view of the respective local node. In each iteration, the best operation is chosen based on the application’s utility functions and a system model that predicts the effects of a given operation. As this optimization process is run on each node independently, it scales well with the number of participants. The prediction only uses local knowledge as well as information from the local neighborhood in up to two hops, which is provided by a neighborhood information exchange protocol. Our evaluation shows that the results of InterestCast’s distributed optimization are close to the global optima computed by a integer program solver. Computing the optimum for a given situation globally at runtime, however, is infeasible due to its computational complexity, even with a highly simplified network model. In detailed network simulations, we further demonstrate the superiority of InterestCast over a purely direct event dissemination in online gaming scenarios. In comparison with the direct dissemination, InterestCast significantly reduces the traffic of weak nodes and almost quadruples the possible number of participants for the same average delivery latency of high-interest events.
|Place of Publication:||Darmstadt|
|Classification DDC:||000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik|
|Divisions:||20 Department of Computer Science
20 Department of Computer Science > Databases and Distributed Systems
|Date Deposited:||02 May 2016 07:45|
|Last Modified:||02 May 2016 07:45|
|Referees:||Buchmann, Ph.D. Alejandro and Nahrstedt, Ph.D. Klara|
|Refereed:||26 February 2016|