RATM: Reinforcement Learning for Co-Optimized CPU Scheduling and NUMA Memory Management in Compiler Design

The current OS uses the stationary heuristic methods & are set-up systematically while designing the system for handling of the CPU schedules & allocating of memories. These heuristics fundamentally fail under the different dynamically shifting workloads characteristic of current data centres, where batch processing, real-time analytics, and interactive applications coexist. This paper introduces RATM (Resource-Aware Adaptive Task Manager), an innovative”Authoritative Controller“architecture developed in Rust that substitutes static policies with a Deep Q-Network (DQN) reinforcement learning agent capable of optimising kernel behaviour during runtime. Our system has a rigorous Policy-Mechanism Separation. The model-free DQN agent watches the system state all the time and chooses actions. The VRRP (Varying Response Ratio Priority) Scheduler and NAAT (NUMA-Aware Adaptive Tiered) Allocator, on the other hand, are passive, configurable mechanisms that carry out commands. The RATM controller acts as a middleman between these layers, making sure that safety rules are followed and turning abstract actions into real API calls. The experimental results show that our RL-driven kernel cuts the average wait time in calibration scenarios by more than 70% compared to the static baseline, while still being very fair. The RL agent learns to proactively activate NUMA page migrations during workload phase changes, which “flattens the curve” of latency spikes that are common in traditional schedulers. Results shows the power of methodology presented here.