Machine Learning-Driven Workload Clustering and Placement for Heterogeneous DRAM/NVM Memory Systems

Abstract

Modern computing systems increasingly face memory bottlenecks, where memory access speeds fail to keep pace with processor advancements. This limitation hinders the performance of memory-intensive applications, necessitating innovative solutions to efficiently manage memory resources. One promising approach is the integration of heterogeneous memory architectures that combine traditional DRAM with emerging non-volatile memory technologies such as Intel Optane. However, leveraging such architectures effectively requires intelligent workload classification and memory allocation strategies. This study explores the application of machine learning techniques to classify workloads based on their memory and computational characteristics. A dataset is constructed by profiling various benchmarks and extracting key performance metrics related to memory bandwidth and CPU utilization. Several supervised learning models, including Random Forest, K-Nearest Neighbors, and Naive Bayes, are trained on this dataset to categorize workloads into predefined classes. Feature selection and dimensionality reduction techniques, such as Principal Component Analysis (PCA), are employed to enhance model performance and interpretability.