Volatility-Aware Hybrid Memory Architecture for Real-Time and Persistent Big Data Systems

Abstract

As data volumes and real-time analytics demands grow, DRAM only memory systems struggle to scale cost-efficiently and sustainably. We present VA-HMA, a volatility-aware hybrid memory architecture that combines DRAM and Non-Volatile Memory (NVM) with a lightweight, machine-learned placement layer and endurance-aware fault tolerance. VA-HMA continuously monitors access patterns and applies a Random Forest–based predictor to guide batched page migrations and adaptive caching, while a durability-aware logging and checkpointing scheme limits NVM wear. Evaluated with DRAMSim3 and NVMain on transactional, analytical, and streaming workloads, VA-HMA achieves up to 35% lower average read latency, up to 25% higher mixed-workload write throughput, 20-30% lower total energy, and 15% reduced NVM write amplification (results reported as mean ± std, n = 5). All simulator configurations, training scripts, and analysis tools are available in the project repository on GitHub. VA-HMA therefore offers a practical, energy-efficient path to scalable, persistent in-memory big-data systems.