Virtual Crowd Interactive Applications

Abstract

This thesis explores virtual crowd interactive applications: applications where large numbers of physically distributed people (tens of thousands) interact as a unified crowd. The purpose is to drive what social psychology identifies as the collective experience, a sense of: unity with many others; belonging; and validation of a shared perspective of the world. Three research objectives are addressed: a design model for collective interaction in a distributed interactive application; a system capable of supporting the interaction requirements of the design model; and a performance test harness for evaluating candidate systems against the interaction requirements of the design model. The first research objective addressed how virtual crowd interactive applications should be designed to drive the collective experience. The model, derived from social psychology, proposes applications where participants: interact over shared state; collaborate on shared input(s); interact directly with a subset of the crowd; and are able to express emotion. From this model two conceptual applications were derived: CrowdALOUD!, where participants submit text in unison to drive crowd chants on a shared visual canvas; and HiveBall, where participants collectively control a player character. The second and third research objectives addressed the architectural challenges of virtual crowd interactive applications. A system was required that could: aggregate large numbers of input events with low latency; support thousands of concurrent users; and broadcast state with low latency. Concentus was developed to meet these requirements, building on concepts from the stream processing domain. To evaluate the performance of Concentus, a test harness was developed to simulate a crowd of users interacting with the system. Concentus and the performance test harness were evaluated with three application load variants: CrowdALOUD! Symbol (low input freq., small data size); CrowdALOUD! Text (low, large); and HiveBall (high, moderate). The supported crowd size was most dependent on the data requirements of applications. Concentus was able to handle 50,000 users for Symbol; 25,000 for Text; and 20,000 for HiveBall. Two aggregation approaches were evaluated: a parallel and a serial one. The serial approach had the best performance for all of the load variants. Finally, with the exception of some HiveBall test runs, the test harness performed as expected.