A Hierarchical Deep Reinforcement Learning Strategy for Self-Adaptive CPU Scheduling in Multi-tenant Database Systems

Multi-tenant database systems face significant challenges in CPU resource allocation due to competing performance requirements across diverse tenant workloads with varying priorities, Service Level Agreements (SLAs), and resource consumption patterns. Traditional static scheduling approaches fail to adapt to dynamic tenant demands and changing system conditions, leading to SLA violations, resource wastage, and degraded overall system performance. This study proposes a Hierarchical Deep Reinforcement Learning (HDRL) strategy for self-adaptive CPU scheduling in multi-tenant database environments. The framework employs a two-tier architecture where a global scheduler manages tenant prioritization and resource allocation policies, while local schedulers optimize CPU utilization within individual tenant contexts. Deep Q-Networks (DQNs) and Proximal Policy Optimization (PPO) algorithms enable dynamic adaptation to evolving tenant workloads and system resource availability. Experimental evaluation using multi-tenant workload benchmarks demonstrates that the proposed strategy achieves 41% improvement in SLA compliance rates while reducing average query response time by 35% across all tenant categories. The hierarchical approach successfully balances resource fairness with performance optimization, resulting in 27% better resource utilization efficiency compared to conventional scheduling methods.