Collaborative Research: CSR: Medium: Towards A Unified Memory-centric Computing System with Cross-layer Support

NSF Award CNS-2310422; $886,226 ($1M collaborative total); October 2023 through September 2026. This project is a collaborative effort with Rujia Wang and Xian-He Sun at IIT and Peng Jiang at the University of Iowa.

Data-centric applications, from scientific simulation to emerging machine learning and data mining algorithms, are becoming prevalent. These applications not only require a large amount of data to compute and store, but also generate massive amounts of intermediate data being moved around the compute resources. Therefore, memory systems have become the bottleneck of computing. Memory-centric computing is a potential solution to overcome the unprecedented memory performance bottleneck. To mitigate bandwidth limitations, computation logic can be added near or in memory chips to reduce data access delay and data movement cost. To overcome memory capacity limitations, the system can incorporate a shared memory pool that all compute units can directly access. Both these memory-centric solutions are promising and machine feasible. However, simply using or combining the two solutions will not realize their full potential. We need to properly integrate the architecture and revisit the full system stack through collaborative designs to make these new technologies more effective for applications and enable more efficient computing. In this proposal, we propose an integrated, full-stack, cross-layer system to enable Unified Memorycentric Computing (UniMCC). We target systems that have near-memory data processors (NDPs) as well as a disaggregated shared memory pool. We work on the entire system stack to utilize these newly proposed hardware solutions from a unified viewpoint of architecture, SW/HW interface, code generation and runtime support, and performance modeling and optimization. We first build an active memory system that supports intelligent data push operations to utilize the NDP hardware. We then revisit existing system software to support features like NDP function calls, data push operations. We also provide a new memory-centric programming model with compiler and runtime optimization for efficient code generation. Lastly, we propose a global performance optimization framework to coordinate all our efforts for the best overall system performance. Putting the four pieces together, our proposed UniMCC system will maximize the potential of NDPs and disaggregated memories and lift memory-centric computing to a new level.

Kyle C. Hale
Kyle C. Hale
Associate Professor of Computer Science

Hale’s research lies at the intersection of operating systems, HPC, parallel computing, computer architecture.