Self-Adaptive Data Processing to Improve SLOs for Dynamic IoT Workloads

摘要

Internet of Things (IoT) covers scenarios of cyber-physical interaction of smart devices with humans and the environment and, such as applications in smart city, smart manufacturing, predictive maintenance, and smart home. Traditional scenarios are quite static in the sense that the amount of supported end nodes, as well as the frequency and volume of observations transmitted, does not change much over time. The paper addresses the challenge of adapting the capacity of the data processing part of IoT pipeline in response to dynamic workloads for centralized IoT scenarios where the quality of user experience matters, e.g., interactivity and media streaming as well as the predictive maintenance for multiple moving vehicles, centralized analytics for wearable devices and smartphones. The self-adaptation mechanism for data processing IoT infrastructure deployed in the cloud is horizontal autoscaling. In this paper we propose augmentations to the computation schemes of data processing component's desired replicas count from the previous work; these augmentations aim to repurpose original sets of metrics to tackle the task of SLO violations minimization for dynamic workloads instead of minimizing the cost of deployment in terms of instance seconds. The cornerstone proposed augmentation that underpins all the other ones is the adaptation of the desired replicas computation scheme to each scaling direction (scale-in and scale-out) separately. All the proposed augmentations were implemented in the standalone self-adaptive agent acting alongside Kubernetes' HPA such that limitations of timely acquisition of the monitoring data for scaling are mitigated. Evaluation and comparison with the previous work show improvement in service level achieved, e.g., latency SLO violations were reduced from 2.87% to 1.70% in case of the forecasted message queue length-based replicas count computation used both for scale-in and scale-out, but at the same time higher cost of the scaled data processor deployment is observed.