NSF Award
2323100
Tiered memory systems are important for companies and organizations that need to manage vast amounts of data while keeping costs under control. They help to improve performance and efficiency by reducing the time it takes to access data, and also reduce the overall cost of computing systems by providing a more cost-effective solution for storing large amounts of data. The research objective of this proposed project is to design new tiered memory management techniques for in-memory databases and analytic frameworks. We will explore the power of machine learning (ML) as well as its limits and overheads as a more general, adaptable solution in improving many aspects such as scanning, migration, allocation, parameter tuning, and task scheduling to optimize the performance, QoS, and resource utilization. This research will allow designing better tiered memory systems for rapidly changing workloads. The proposed comprehensive approach to applying ML to tiered memories will generate new guidelines and leave a significant impact on many areas that are dependent on processing a large amount of data. This project will share the findings with undergraduate and graduate students through computer science programs and open up career opportunities to students from underrepresented groups and first-generation college students. This project will disseminate the proposed techniques into the industry and foster technology transfer through new industrial collaborations. The developed infrastructure will be available to the research community through a web-based portal.<br/><br/>This research makes empirical contributions to the system-ML co-design space by tackling major challenges posed by evolving large-scale memory-intensive applications. Specifically, it advances the state of knowledge regarding, (1) how to design and develop a machine learning-based dynamic memory tiering system that is designed to ensure that the right data is in the right tier at the right time?; (2) how to use expert domain knowledge for each ML design decision such that the efficiency and the overhead of the final model are manageable and useful in real systems?; (3) how to design interactive frameworks that allow the user to modify the internal resources and parameters of the tiered memory system?; (4) how to enable novices to configure tiered memory systems with respect to their workloads and data processing requirements to obtain high performance and resource utilization?; and (5) how to derive ML models to predict the future workload patterns and accordingly configure the tiered memory systems in advance for better performance?; and Thus, to design sustainable ML-based tiered memory systems, exploring the trade-off between performance, QoS, and resource utilization by using different ML models for various design decisions is very important.<br/><br/>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.
