CRII: CSR: Programmable Heterogeneous Memory Systems via Multiple Address Spaces and RAM Lake

点击翻译按钮获取中文摘要

Computer systems store an increasingly vast amount of data. The major limiting factor for computer performance is data retrieval from memory, not data processing. At the same time, these systems integrate various physical storage technologies adding complexity to data retrieval. The project will develop new improvements to computing system hardware to increase data retrieval performance in systems with many memory technologies by addressing two initial questions: how to move data between memory technologies efficiently and how to represent different types of data flexibly for both the programmer and the hardware system.This project will produce new hardware mechanisms and software interfaces that will increase the performance and simplify the programming of heterogeneous memory systems. Additionally, the work will expand the understanding of how software will use these nascent heterogeneous memory systems by deeply studying current and emerging systems and applications. This work will be broken into two main thrusts. First, designing novel hardware support for fine-grained user-directed and system-directed data movement, and second, designing hardware support for multiple address spaces to improve the interfaces to heterogeneous memory for programmers and system software.Computing systems ranging from datacenters and supercomputers to mobile devices and Internet-of-Things systems are processing an increasing amount of data, which is driving the trend toward heterogeneous memory systems. The proposed research will increase the performance and usability of these computing systems. This research will be presented to the scientific community and the public through conference publications, software repositories, and blog posts on a public website. Additionally, this project will develop an upper-divisional undergraduate/introductory graduate class on programming heterogeneous systems.All data, software, and other artifacts generated through this research will be carefully managed and archived through third party repositories such as GitHub and cloud service providers. The software artifacts will be maintained indefinitely using publicly accessible history tracking version control systems. Data that can be regenerated by re-running simulation software will be kept for at least five years, as well as the foreseeable future. Links to repositories containing the simulators used, code produced, and data used in the published results can be found at https://arch.cs.ucdavis.edu/projects/heterogeneous-memory.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.