CAREER: Towards Efficient Accelerated Cloud Data Centers

职业：迈向高效加速云数据中心

• 批准号: 2047521
• 负责人: Daniel Wong
• 金额: $ 51.07万
• 依托单位: University of California-Riverside
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-03-01 至 2026-02-28
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2047521&HistoricalAwards=false
• 关键词: CAREER; Towards; Efficient; Accelerated; Cloud

Emerging large-scale compute-intensive applications, such as machine learning and big data analytics, have led to the wide-spread adoption of compute accelerators, such as Graphical Processing Units (GPUs), in cloud data centers. However, the existing cloud management software stack introduces many layers of abstraction that strip away application characteristics and hardware architectural details leading to inefficient and uncoordinated cloud management policies. This can lead to slow down of application performance and under-utilization of hardware resources, which ultimately impacts the data center's total cost of ownership. This project will design efficient accelerated cloud data centers that are performance-efficient, resource-efficient, and cost-efficient. The planned research has three main goals: (1) develop software frameworks to measure and identify the causes of inefficiencies in accelerated cloud data centers, (2) design inter-accelerator communication-aware cloud management policies, and (3) design accelerator-assisted interconnect topologies.The success of this project can improve application performance, improve hardware resource utilization, and reduce energy consumption; leading to greener data centers and reduced carbon footprint. In addition, this project will enable data centers to cost-efficiently scale computational power to keep pace with the growing societal demands of emerging machine learning and artificial intelligence applications. The tools and frameworks developed in this project will be made publicly available to facilitate the efficient integration of compute accelerators in cloud data centers. These new tools and frameworks will form the foundation for new course development, undergraduate research opportunities, and outreach efforts to train and prepare a new generation of engineers who will utilize compute accelerators and cloud resources as a first-class design choice.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.

新兴的大规模计算密集型应用，如机器学习和大数据分析，导致云数据中心广泛采用计算加速器，如图形处理单元(GPU)。然而，现有的云管理软件堆栈引入了许多抽象层，剥离了应用程序特征和硬件架构细节，导致了低效和不协调的云管理策略。这可能会导致应用程序性能下降和硬件资源利用不足，最终影响数据中心的总拥有成本。该项目将设计高效的加速云数据中心，具有性能效率、资源效率和成本效益。计划中的研究有三个主要目标：(1)开发软件框架来测量和识别加速云数据中心效率低下的原因，(2)设计加速器间通信感知的云管理策略，(3)设计加速器辅助的互联拓扑。该项目的成功可以提高应用程序性能，提高硬件资源利用率，降低能耗；导致更绿色的数据中心和减少碳足迹。此外，该项目将使数据中心能够经济高效地扩展计算能力，以跟上新兴机器学习和人工智能应用日益增长的社会需求。该项目中开发的工具和框架将公开提供，以促进云数据中心中计算加速器的高效集成。这些新的工具和框架将为新课程开发、本科研究机会以及培养和培养新一代工程师的外展工作奠定基础，这些工程师将把计算加速器和云资源作为一流的设计选择。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。

项目成果

WattWiser: Power & Resource-Efficient Scheduling for Multi-Model Multi-GPU Inference Servers

WattWiser：电源

• DOI: 10.1145/3634769.3634807
• 发表时间: 2023
• 期刊: Proceedings of the 14th International Green and Sustainable Computing Conference
• 作者: Ali Jahanshahi;Mohammadreza Rezvani;Daniel Wong
• 通讯作者: Daniel Wong

MAPA: Multi-Accelerator Pattern Allocation Policy for Multi-Tenant GPU Servers

• DOI: 10.1145/3458817.3480853
• 发表时间: 2021-10
• 期刊: SC21: International Conference for High Performance Computing, Networking, Storage and Analysis
• 作者: K. Ranganath;Joshua D. Suetterlein;J. Manzano;S. Song;Daniel Wong