Joint Operator Scaling and Placement for Distributed Stream Processing Applications in Edge Computing

Abstract

Distributed Stream Processing (DSP) systems are well acknowledged to be potent in processing huge volume of real-time stream data with low latency and high throughput. Recently, the edge computing paradigm shows great potentials in supporting and boosting the DSP applications, especially the time-critical and latency-sensitive ones, over the Internet of Things (IoT) or mobile devices by means of offloading the computation from remote cloud to edge servers for further reduced communication latencies. Nevertheless, various challenges, especially the joint operator scaling and placement, are yet to be properly explored and addressed. Traditional efforts in this direction usually assume that the data-flow graph of a DSP application is pre-given and static. The resulting models and methods can thus be ineffective and show bad user-perceived quality-of-service (QoS) when dealing with real-world scenarios with reconfigurable data-flow graphs and scalable operator placement. In contrast, in this paper, we consider that the data-flow graphs are configurable and hence propose the joint operator scaling and placement problem. To address this problem, we first build a queuing-network-based QoS estimation model, then formulate the problem into an integer-programming one, and finally propose a two-stage approach for finding the near-optimal solution. Experiments based on real-world DSP test cases show that our method achieves higher cost effectiveness than traditional ones while meeting the user-defined QoS constraints.