ASTRA: Adversarial Sim-to-Real Transfer Reinforcement Learning for Autoscaling in Cloud Systems

With the widespread adoption of cloud computing, autoscaling has become crucial for efficient resource management and stable service provision in cloud systems. In recent years, autoscaling methods based on deep reinforcement learning (DRL) have gained significant attention due to their outstanding adaptability and flexibility. However, training DRL-based autoscaler requires interactions with real cloud systems, incurring high interaction costs, low data collection efficiency, and potential operational impacts. To address these challenges, we propose ASTRA, a sim-to-real transfer reinforcement learning framework for autoscaling. ASTRA constructs a cloud system simulation environment based on a performance estimation model, enabling low-cost and high-efficiency training sample collection for policy learning. The learned policy is subsequently transferred to the real systems for scaling decisions. To address performance modeling inaccuracies caused by dynamic cloud state changes, we propose a performance modeling method based on hybrid attentive state space model. By incorporating state space model, it captures system dynamics and state evolution, effectively reducing simulation errors. Furthermore, to mitigate the performance degradation of the transferred policy due to the distribution shift, we propose an autoscaling method based on adversarial soft actor-critic. By introducing adversarial policy training with gradient regularization based on state perturbations, it significantly improves transferred policy performance. The results in the real system demonstrate that ASTRA achieves optimal overall performance in environment modeling, policy transfer and real-world autoscaling. Specifically, ASTRA outperforms all baselines in terms of instance number, response time, SLO violation rate, and CPU utilization under different workload patterns. More importantly, under limited interaction costs, ASTRA achieves a 616.94× improvement in interaction sample collection rate compared to direct online training method.