CAREER: Rethinking Rack-Scale System Design over Memory Fabrics
- Institution
- University of Wisconsin-Madison, WI
- Principal investigator
- Ming Liu
- Funding
- $569K
- Duration
- 2024–2029
Moonbase score
Still being scored
Abstract
Memory fabrics, an emerging and powerful cluster interconnect technology, are revolutionizing the construction of the modern data center. Albeit promising, this infrastructure shift fundamentally changes decade-long system design assumptions and challenges conventional wisdom on how to build efficient rack-scale systems. This project aims to develop a new computing paradigm that views the memory fabric as a first-class citizen on which to instantiate, orchestrate, and reclaim computations. The project’s novelties are that a memory fabric-aware intermediate system stack allows applications to effectively harness the capabilities of memory fabrics and hide their limitations. The project’s broader significance and importance are laying out the system foundation for building next-generation sustainable and cost-efficient computing infrastructures for enterprise on-premise and cloud-scale data center systems using memory fabrics. This project introduces the technology to a broader audience via the LegoCluster program and BigComputer workshop education activities. This project rethinks the rack-scale system design from memory, computing, and communication perspectives. The first thrust classifies remote memory nodes into four categories and develops a benchmarking framework to characterize their performance. It builds an active remote memory system that transparently and dynamically moves a memory object to a suitable memory node at runtime based on the object’s access profile and the underlying node’s capabilities. The second thrust applies idempotent tasks as the fault-tolerant computation abstraction and develops a language system to generate them. It designs hardware cooperative engines and an execution framework on which to schedule idempotent tasks. The third thrust introduces new communication primitives, translates them into the memory fabric’s commands, and builds a transport layer for end-to-end traffic control. Together, these efforts produce a new open-source software system over memory fabrics encompassing APIs, libraries, programming systems, and software runtimes. 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.
LOADING MOONBASE SCORE
Questions for University of Wisconsin-Madison
Answered by the team behind this research.
No questions yet. Sign in to ask the first one
Related grants
Same researchers, same institution, same topic, scored first.
- NSF · University of California-RiversideNot yet scored
Engineering Nanoscale Light-Heat Interactions in 2D Devices: Resolving Photovoltaic/Photothermoelectric Competition for High-Performance Optoelectronics
Ming Liu · $400K · 2025–2028
- NSF · University of UtahNot yet scored
CloudLab Phase-IV: Strengthening a Core Resource for Research and Education
Robert Ricci · $12.0M · 2024–2028
- NSF · University of Wisconsin-Madison69
Next-Generation Safe and Dexterous Robot Hands using High-Torque Direct-Drive Actuation
Lei Zhou · $566K · 2023–2026