CSR: Small: Processing-in-Memory enabled Manycore Systems to Accelerate Graph Neural Network-based Data Analytics

许多科学和工程应用都是通过处理大型和异类的图结构数据来实现的。例如，来自传感器馈送、有关事件的信息数据库、供应链和网络流量的数据本质上是关系的，需要基于图形的表示。图形神经网络(GNN)将允许分析从数据中发现隐藏的模式，并使这些应用程序能够进行预测和决策支持。边缘训练机器学习(ML)模型(片上训练或嵌入式系统训练)可以解决包括数据隐私/安全在内的许多紧迫挑战，通过减少对通信结构和云基础设施的依赖来增加ML应用程序在世界不同地区的可访问性，并满足增强/虚拟现实(AR/VR)应用的实时要求。包括AR/VR在内的许多应用都需要在嵌入式系统上进行GNN培训。然而，现有的EDGE平台不具备足够的能力来支持GNN的现场培训。此外，据估计，在传统计算平台(如GPU)上训练一个未经修剪的神经网络的成本可能超过1万美元，在其生命周期内排放的碳相当于五辆汽车的排放量。基于电阻随机存取存储器(ReRAM)的内存中处理(PIM)架构是解决这一问题的一种很有前途的解决方案。基于ReRAM的体系结构的交叉开关结构实现了高效的矩阵向量乘法(MVM)运算，这在包括GNN训练/推理在内的现代ML任务中普遍存在。我们以ReRAM为例，但所提出的计算框架同样适用于任何其他基于纵横制的PIM配置。这项工作的教育贡献在于建立了一个基于跨学科研究的课程，整合了PIM、机器学习和数据驱动的设计优化。拟议的研究将通过使学生能够应用课堂知识来研究需要硬件、软件和理论专业知识的问题，从而加强对学生的教育。私人投资机构在让代表性不足的群体参与研究方面积累了多年的经验。这一经验将被用来激励和吸引来自代表性不足群体的学生，包括妇女、非裔美国人和西班牙裔美国人。在这个项目中，我们为使用基于PIM的多核系统的GNN计算建立了一个新颖而可靠的计算框架。随着基于GNN的应用从边缘到云的需求不断增长，我们需要计算系统来满足严格的大小、重量和功率(交换)限制。这项研究的主要贡献将是高性能、高能效和可靠的基于PIM的GNN计算体系结构的概念开发、优化和评估。尽管对基于GNN的数据分析的兴趣和广泛的研究和应用研究呈指数级增长，但硬件辅助执行效率和硬件感知算法效率的重要性并没有得到足够的重视。这项研究将减少执行基于GNN的应用程序时对数据中心和高性能计算(HPC)集群的依赖。目前还缺乏全面的解决方案，使我们能够为GNN快速设计和优化支持PIM的计算平台。因此，本文提出的支持机器学习的软硬件协同设计优化策略将对越来越多地部署GNN的计算平台产生深远的影响。越来越多的EDGE ML应用程序通过在移动平台或嵌入式系统上部署GNN来实现。由于移动/嵌入式平台同时受到计算和存储的限制，迫切需要基于PIM的解决方案来为EDGE ML应用部署GNN。更多的边缘计算能力将减少互联网流量和数据中心的耗电处理。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Many science and engineering applications are enabled by processing large and heterogeneous graph structured data. For example, data from sensor feeds, information databases about events, supply chains, and web traffic are relational by nature and require graph-based representations. Graph Neural Networks (GNNs) will allow the analysis to uncover hidden patterns from the data and can enable these applications for making predictions and decision support. Training machine learning (ML) models at the edge (training on-chip or on embedded systems) can address many pressing challenges, including data privacy/security, increase the accessibility of ML applications to different parts of the world by reducing the dependence on the communication fabric and the cloud infrastructure, and meet the real-time requirements of augmented/virtual reality (AR/VR) applications. Many applications including AR/VR require GNN training on embedded systems. However, existing edge platforms do not have sufficient capabilities to support on-device training of GNNs. Moreover, it is estimated that training a single unpruned neural network on conventional compute platforms, such as GPUs, can cost over $10,000 and emit as much carbon as five cars over their lifetimes. Resistive random-access memory (ReRAM) based processing-in-memory (PIM) architectures are a promising solution to address this problem. The crossbar structure of ReRAM-based architectures enables efficient Matrix-Vector Multiplication (MVM) operations, which are ubiquitous in modern ML tasks including GNN training/inference. We use ReRAM as an example, but the proposed computing framework will work equally well for any other crossbar-based PIM configuration. The educational contribution of this work lies in the establishment of an interdisciplinary research-based curriculum integrating PIM, machine learning, and data-driven design optimization. The proposed research will enhance the education of students by enabling them to apply classroom knowledge to research problems that require hardware, software, and theoretical expertise. The PIs have many years of accumulated experience in involving underrepresented groups in research. This experience will be leveraged to motivate and engage students from underrepresented groups, including women, African Americans, and Hispanics. In this project, we lay the foundations for a novel and reliable computing framework for GNN computation using PIM-based manycore systems. With the rising needs of GNN-based applications from the edge to the cloud, we need computing systems to meet the stringent size, weight, and power (SWaP) constraints. The key contribution of this research will be the conceptual development, optimization, and evaluation of high-performance, energy-efficient, and reliable PIM-based architectures for GNN computing. Despite the exponential growth in interest and extensive research and application studies on GNN-based data analytics, the importance of hardware-assisted execution efficiency and hardware-aware algorithm efficiency has not received adequate attention. This research will reduce the dependency on data centers and high-performance computing (HPC) clusters for executing GNN-based applications. There is a lack of holistic solutions that allow us to quickly design and optimize PIM-enabled computing platforms for GNNs. Hence, machine learning enabled hardware and software co-design optimization strategies proposed in this work will have profound impacts on computing platforms where GNNs are increasingly deployed. There is a growing set of edge ML applications that are enabled by deploying GNNs on mobile platforms or embedded systems. Since mobile/embedded platforms are constrained by both compute and storage, there is a great need for PIM-based solutions to deploy GNNs for edge ML applications. More computational power at the edge will reduce both the internet traffic and power-hungry processing at data centers.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.