Enhanced Hardware-Aware Memory Allocation Using Reinforcement Learning and Spectral Graph Analysis for Heterogeneous Embedded Systems

**Abstract:** This paper presents a novel memory allocation framework, HARMA (Hardware-Aware Reinforcement Memory Allocator), specifically designed for heterogeneous embedded systems characterized by diversified memory technologies (e.g., DRAM, SRAM, eDRAM) each offering different performance and power trade-offs. HARMA dynamically optimizes allocation decisions by integrating reinforcement learning (RL) with spectral graph analysis of memory architecture and application workload properties. Unlike traditional static and heuristic-based allocators, HARMA learns through interaction with the hardware environment, resulting in a 15-30% improvement in overall system performance and a 10-20% reduction in energy consumption in simulated environments comprising ARM Cortex-A72 cores, integrated GPUs, and varying memory configurations. This framework directly addresses the limitations of existing deterministic allocators that fail to account for the hardware heterogeneity and workload characteristics of modern embedded platforms.