EAGER: Intelligent Memory Management

About This Grant

This project addresses the complexities of managing memory efficiently in modern heterogeneous computing environments, which often involve a mix of various types of processor and memory. Efficient memory management must deal with the high costs of transferring memory between multiple memory types. Requesting this data can dramatically increase access latency, so there is a strong motivation to identify and preemptively start transferring data before it is needed. Current memory management approaches often struggle to handle rapidly changing workloads, while scaling to systems with many types of memory and ensuring robust security. Inefficient memory management can hinder performance, energy efficiency, and security in edge computing and large data centers. The project’s novelties are its adaptive memory management that uses real-time machine learning model training for fast, accurate predictions; its ability to support a wide array of memory types; and its built-in security features for fault protection and data consistency. The project's impacts are broad, with applications in edge computing, autonomous systems, and cloud platforms, applying to a wide variety of memory types. Ultimately, this research aims to make these complex computing systems faster and easier for development while also reducing energy consumption in data centers. The resulting software products will be released as an open source release, to foster widespread adoption. The technical approach of this project centers on an intelligent memory management framework that integrates predictive analysis, dynamic resource reallocation, and unified memory security. The investigator will employ machine learning models, including Long Short-Term Memory (LSTM) networks, Reinforcement Learning (RL), and Deep Neural Networks (DNN), to predict memory access patterns in real time. These predictions will guide a novel multi-tier-aware scheduler to dynamically balance memory loads across diverse memory locations such as DRAM, NVM, and HBM. Security is a key component, enforced through compile-time instrumentation and runtime consistency checks, generalizing across various memory types. Expected advances include the development of an adaptive memory management system capable of significantly reducing retraining times for its machine learning models, from hours to minutes, and optimizing heterogeneous memory through an ML-aided runtime management paradigm, intelligently and preemptively moving data where it is needed to reduce data transfer latency. The project will also focus on generalizing memory management to support multiple types of memory and incorporating security for fault protection and consistency. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.