A Workload Analysis of Live Event Broadcast Service in Cloud

Abstract

The class of popular Internet live events, such as sports game broadcast services, can greatly benefit from cloud technology that employs resource pooling and rapid elasticity in resource allocation and management. The external demands tend to be unpredictable and exhibit a high degree of burstability. To sustain a good viewing experience for the users, an important QoS support for this class of applications is the peak load management in the presence of unpredictable demand. This relies on a close grasp of the demand behavior characteristics and an accurate prediction model for them. In this paper, we analyze live sports event broadcast service workloads from a commercial Internet service provider. Our results show that popularity follows a 2-mode Zipf distribution. We also observe some significant characteristics of the demand behavior. First, the demand behavior may differ significantly between games. Popular events tend to exhibit highly variable behaviors in time, volume and the change rate during the course. Finally, the demand variation highly correlates with certain event-specific time points. We conclude the paper with a preliminary study that applies three simple statistical models in workload prediction at runtime as an input to dynamic resource resizing in a cloud. The results indicate that more e ective ways are needed to better capture the dynamics and unpredictability of the workload to improve prediction accuracy.