Online resource provisioning of applications in cloud is challenging due to the variable nature of workloads and the interference caused by sharing resources. Current on-demand scaling is based on manually configured thresholds that cannot capture the dynamics of applications and virtual infrastructure. This results in slow responses or inaccurate provisioning that lead to unfulfilled service level objectives (SLOs). More automated approaches, in turn, use fixed model structures and feedback loops to control key performance indicators (KPIs). However, workload surges and the non-linear behavior of software (e.g. overload control) make the control mechanisms vulnerable to rapid variations, eventually leading to oscillatory or unstable elasticity. In this paper we introduce RobOps, a robust control system for automated resource provisioning in cloud. RobOps incorporates online system identification (SID) to dynamically model the application and detect variations in the underlying hardware/software. Our framework combines feedforward/feedback control with prompt response to achieve reference performance. The feedforward control allows to compensate for delays in the scaUng mechanism and provides robustness to workload surges. We validate RobOps performance using an enterprise communication service. Compared to baseline approaches, RobOps achieves 2X less SLO violations in case of traffic surges, and reduces the impact of interferences at least by 20%.
展开▼
