EAGER: Adaptive Shared Memory Management for Heterogeneous CPU-GPU Architectures

Virginia Commonwealth University, Richmond VA

Investigators

Abstract

Building efficient memory systems for exascale computing has nowadays become extremely challenging. Applications running in exascale computing environments are becoming increasingly diverse and complicated, making managing shared memory system particularly difficult. Graphic Processing Units (GPUs) are entering exascale computing as accelerators, imposing new technical challenges in sharing memory. Non-volatile memory (NVM) has promising future in the memory/storage hierarchy. It can be used to enlarge memory capacity and/or improve energy efficiency. However, how to efficiently integrate NVM into current memory architecture to build a highly efficient memory system remains an open problem. This project proposes an adaptive shared memory management scheme to address the memory interference problem in heterogeneous CPU-GPU architectures. A flexible framework based on a hybrid memory model for integrating non-volatile memory into a shared memory system will be developed. In addition, the project will explore designs to build a unified address space for heterogeneous architectures. This project addresses the challenges in exploring shared memory management schemes for heterogeneous architectures. Proposed techniques of management for memory interference in both homogeneous and heterogeneous architectures are well amenable to the high concurrency and widely different interference patterns in exascale computing environments. These techniques provide a compelling solution to managing shared memory for heterogeneous systems. The proposed work can serve as a starting point to conquer the bigdata and exascale computing challenges in terms of large-scale memory and computer system design. Moreover, the proposed research will facilitate the server clusters for memory-intensive applications to most effectively utilize those existing/emerging architecture and system technologies to tackle the memory challenge in heterogeneous CPU-GPU systems. The proposed efficient shared memory management can benefit numerous memory-intensive applications such as big data analytics, biology, chemistry, earth science, etc. In addition, the project will integrate research and education together. It will provide opportunities for undergraduate and graduate students to participate in the research and help train a new generation of computer scientists and engineers in the area of high-performance computing.