Spike-based Dynamic Data Placement over Hybrid DRAM/NVM Memory Systems

Abstract

Modern computing systems require more and more high-capacity memory solutions that deliver both high performance and sustainability. To address these needs, heterogeneous memory architectures—integrating traditional DRAM with Non-Volatile Memory (NVM) technologies—are emerging as effective alternatives. However, these systems introduce complex challenges in managing dynamic data placement between DRAM and NVM. In this work, we introduce SPID, a lightweight, spike-based data placement algorithm for heterogeneous DRAM/NVM systems. SPID leverages advanced bandwidth and active object spike detection mechanisms and adaptive decision-making to optimize memory utilization efficiently. We evaluate SPID against a set of baseline placement policies and state-of-the art solutions over real-life applications, showing that we achieve 30.82% higher performance and 31.61% less energy consumption on average. SPID is integrated in SPMalloc, an open-source library that intercepts dynamic memory allocation calls at runtime, enabling precise monitoring of allocation patterns with minimal overhead. Together, SPID and SPMalloc optimize data placement, enhancing performance in heterogeneous memory environments.