Single-Input Multiple-Output Control for Multi-Goal Orchestration

In this paper, we propose a QoS-aware Single Input Multiple Output (SIMO) controller that combines performance and cost goals while aiming to maintain system stability. To enhance robustness, as targeted by inspiring control concepts, we use system identification models and analytical tuning techniques for Proportional-Integral-Derivative (PID) controllers. Our resulting SIMO PI controller performs well when tracking reference values that may change over time and when conciliating conflicting goals according to the user's preference. In contrast, a naïve use of independent controllers may lead to opposing decisions and instabilities, as the controllers work against each other. We examine the use of the controller to orchestrate processing pods in a Kubernetes cluster for an IoT sensor analysis application (power consumption disaggregation). Nevertheless, the lessons learned in the design of the controller apply to other use cases, including batch and interactive workloads.