CAREER: Automated and Efficient Machine Learning as a Service

职业：自动化高效的机器学习即服务

• 批准号: 2048044
• 负责人: Feng Yan
• 金额: $ 51.75万
• 依托单位: Board of Regents, NSHE, obo University of Nevada, Reno
• 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: 2021-05-01 至 2023-05-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2048044&HistoricalAwards=false
• 关键词: CAREER; Automated; Efficient; Machine; Learning

Machine-Learning-as-a-Service (MLaaS) is an emerging computing paradigm that provides optimized execution of machine learning tasks, such as model design, model training, and model serving, on cloud infrastructure. Explosive growth in model complexity and data size along with the surging demands of MLaaS is already resulting in substantial increases in computational resource and energy requirements. Unfortunately, existing MLaaS systems have poor resource management and limited support for user specified performance and cost requirements, exacerbating waste in computing resources and energy. This project aims to utilize the unique features of MLaaS to design efficient, automated, and user-centric MLaaS systems. This approach will significantly reduce resource waste and shorten the model design cycles through a variety of novel optimization approaches and by eliminating candidate models that fail to meet model serving latency and target accuracy. To support complete MLaaS workflow, this project will also develop MLaaS model serving methodologies that can meet service level latency requirements with minimum resource consumption using intelligent autoscaling.This project has the potential to tremendously reduce the resource and energy consumptions as well as the carbon footprint associated with the fast-growing societal demands in machine learning and cloud computing. Important insights and technologies will be produced targeting resource management and energy saving of the next-generation machine learning systems and cloud infrastructure. The findings of this project will also contribute to related fields of parallel and distributed systems, performance evaluation and optimization, and green computing. This project will carry out substantial integrated education activities including new course and online education development, integration of industry feedback in education. Additionally, the work will impact undergraduate and graduate students by training them in the art of system optimization combined with the latest machine learning domain knowledge while combining outreach and engagement of students from underrepresented groups and especially women.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.

机器学习即服务（MLaaS）是一种新兴的计算范式，它在云基础设施上提供机器学习任务的优化执行，例如模型设计，模型训练和模型服务。模型复杂性和数据大小的爆炸式增长沿着MLaaS需求的激增，已经导致计算资源和能源需求的大幅增加。不幸的是，现有的MLaaS系统具有较差的资源管理和对用户指定的性能和成本要求的有限支持，从而加剧了计算资源和能量的浪费。该项目旨在利用MLaaS的独特功能来设计高效，自动化和以用户为中心的MLaaS系统。这种方法将通过各种新颖的优化方法以及通过消除未能满足模型服务延迟和目标准确性的候选模型来显著减少资源浪费并缩短模型设计周期。为了支持完整的MLaaS工作流程，该项目还将开发MLaaS模型服务方法，该方法可以使用智能自动缩放以最小的资源消耗来满足服务级别延迟要求。该项目有可能极大地减少资源和能源消耗以及与机器学习和云计算快速增长的社会需求相关的碳足迹。将针对下一代机器学习系统和云基础设施的资源管理和节能产生重要的见解和技术。本计画之研究成果亦将有助于平行与分散式系统、效能评估与最佳化、绿色计算等相关领域。该项目将开展包括新课程和在线教育开发、整合教育行业反馈等在内的实质性整合教育活动。 此外，这项工作将通过培训本科生和研究生掌握系统优化的艺术，结合最新的机器学习领域知识，同时结合对代表性不足的群体，特别是女性学生的拓展和参与，对他们产生影响。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值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

DySR: Adaptive Super-Resolution via Algorithm and System Co-design

DySR：通过算法和系统协同设计实现自适应超分辨率

• 发表时间: 2023
• 期刊: International Conference on Learning Representations (ICLR 2023
• 作者: Zawad, Syed;Li, Cheng;Yao, Zhewei;Zheng, Elton;He, Yuxiong;Yan, Feng
• 通讯作者: Yan, Feng

AUTOGR Automated Geo-Replication with Fast System Performance and Preserved Application Semantics

AUTOGR 自动异地复制，具有快速的系统性能和保留的应用程序语义

• DOI: 10.14778/3461535.3461541
• 发表时间: 2021
• 期刊: Proceedings of the VLDB Endowment
• 影响因子: 2.5
• 作者: Wang Jiawei;Li Cheng;Ma Kai;Huo Jingze;Yan Feng;Feng Xinyu;Xu Yinlong
• 通讯作者: Xu Yinlong

SimiGrad: Fine-Grained Adaptive Batching for Large Scale Training using Gradient Similarity Measurement

• 发表时间: 2021
• 作者: Heyang Qin;Samyam Rajbhandari;Olatunji Ruwase;Feng Yan;Lei Yang;Yuxiong He

lunule: An Agile and Judicious Metadata Load Balancer for CephFS

• DOI: 10.1145/3458817.3476196
• 发表时间: 2021-11
• 期刊: SC21: International Conference for High Performance Computing, Networking, Storage and Analysis
• 作者: Yiduo Wang;Cheng Li;Xinyang Shao;Youxu Chen;Feng Yan;Yinlong Xu

Optimizing Inference Serving on Serverless Platforms

• DOI: 10.14778/3547305.3547313
• 发表时间: 2022-06
• 期刊: Proc. VLDB Endow.
• 作者: Ahsan Ali;Riccardo Pinciroli;Feng Yan;E. Smirni