Collaborative Research: CNS Core: Medium: Terabyte-scale Tiered Memory Management

点击翻译按钮获取中文摘要

As application demand for memory increases at an explosive pace, we witness a slowdown in the rate at which the currently dominant computer memory technology ("DRAM") scales up in capacity. This growing gap leads to a need for large-capacity memory systems that either disaggregate the DRAM memory components across a network, or adopt a memory technology that offers higher capacity than DRAM but at slower performance. This trend toward larger memories that are split into different performance tiers within a single compute node poses challenges that current memory management mechanisms and techniques cannot keep up with. The critical challenges include how to characterize the memory behavior of applications and workloads and how, when, and where to place and migrate data with low performance, energy, and monetary overhead. This research project will explore these fundamental challenges and develop and evaluate solutions specifically for these large-scale, tiered memory systems. To be most effective, these solutions will span both the computer hardware (i.e., the processor and memory modules) and system software (i.e., the operating system). The combined hardware-software research approach, along with continuous prototyping of the proposed solutions, ensures that the challenges targeted are real and that the solutions will have impact not only on academia, but also on industry and end users. This research is timely and necessary because a comprehensive, low-overhead, tiered memory management system is a prerequisite for unleashing the potential of emerging memory technologies. These technologies are, in turn, necessary, especially for cloud computing, to achieve the performance levels needed for applications, while keeping costs, both monetary and environmental, low. First, the effective use of tiered memories will both reduce the number of memory components installed in systems, thus reducing the embedded and operational carbon associated with them. Second, the developed management techniques will enable a single computing node to successfully serve a higher application load, reducing the carbon footprint of compute as well as memory. Other societal benefits include the unique training this hardware-software research project will provide to students, including undergraduate and graduate students. The project also has a high likelihood of broadening participation in computing. The primary investigators on this project have a track record of advising female students; the participating universities are committed to broadening participation; and the student recruitment environment benefits from a large number of students from groups that are historically underrepresented in computer technology. The University of Texas at Austin is a recognized Hispanic-Serving University.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.