Dynamic Resources Allocation among Collocated Applications via Reinforcement Learning

Cloud servers usually face low utilization. One method to mitigate such problem is to collocate multiple applications onto a single server to improve the resources exploitation. However, due to the workload variations and resources constrains, collocating applications interfere with each other, and this incurs performance degradation. The lack of effective resources managing mechanism further drops the utilization as the number of latency-sensitive user-facing applications, such as search engines and online shopping, increases. In this paper, we assume the strategy is to collocate one latency-critical task with one or more batch tasks. We design a reinforcement learning based controller that can simultaneously meet the performance target of the latency-critical task and improve the throughput of batch tasks. We combine the concept of workload characterization to make our design non-intrusive and widely applicable to any applications. The online reinforcement learning helps with capturing workload variations. We implement a prototype of our design and evaluate on real-world applications. The result indicates its effectiveness compared with uncontrolled collocation and the static feedback-based controller.