多GPU系统非一致存储访问优化:研究进展与展望  

Non-Uniform Memory Access Optimization on Multi-GPU Systems:Research Progress and Prospect

作　　者：李晨[1] 刘畅 葛一漩 郭阳[1] LI Chen;LIU Chang;GE Yi-xuan;GUO Yang(College of Computer,National University of Defense Technology,Changsha,Hunan 410073,China)

机构地区：[1]国防科技大学计算机学院,湖南长沙410073

出　　处：《电子学报》2024年第5期1783-1800,共18页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62202478);国防科技大学自主创新科学基金(No.23-ZZCX-JDZ-12)。

摘　　要：随着晶体管缩小速度的减缓,单GPU(Graphics Processing Units)的性能提升已经变得越来越具有挑战性,因此,多GPU系统成为了提高GPU系统性能的主要手段.然而,由于片外物理设计的制约,多GPU系统中处理器间的带宽不均衡导致了非一致存储访问(Non-Uniform Memory Access,NUMA)问题,严重影响多GPU系统的性能.为了减少非一致存储访问所导致的性能损失,本文首先分析了非一致存储访问出现的原因,并对现有的非一致存储访问解决方案进行了对比.针对不同维度的非一致存储访问,本文从减少远程访问流量和提升远程访问性能两个方向出发,对非一致存储访问的优化方案进行了总结.最后,结合这些方案的优缺点,提出了未来多GPU系统非一致存储访问优化的发展方向.Due to the slowdown of transistor scaling,it has become increasingly difficult to enhance the performance of a single GPU(Graphics Processing Units).Therefore,multi-GPU systems have become the main means to improve the performance of GPU systems.However,due to the constraints of off-chip physical design,the bandwidth imbalance between processors in multi-GPU systems leads to non-uniform memory access(NUMA)problems,which seriously affects the performance of multi-GPU systems.In order to reduce the performance loss caused by non-uniform memory access,this paper first analyzes the causes of non-uniform memory access and compares existing solutions for non-uniform memory access.For non-uniform memory access with different dimensions,this paper summarizes optimization solutions for non-uniform memory access from two directions:reducing remote access traffic and improving remote access performance.Finally,combining the advantages and disadvantages of these solutions,this paper proposes the future development direction of nonuniform memory access optimization for multi-GPU systems.

关 键 词：多GPU系统 非一致存储访问 GPU访存 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]