Collaborative Research: PPoSS: Planning: Integrated Scalable Platform for Privacy-aware Collaborative Learning and Inference

协作研究：PPoSS：规划：用于隐私意识协作学习和推理的集成可扩展平台

• 批准号: 2029004
• 负责人: Vivek Sarkar
• 金额: $ 15万
• 依托单位: Georgia Tech Research Corporation
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2023-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2029004&HistoricalAwards=false
• 关键词: Collaborative; Research; PPoSS; Planning; Integrated

Building scalable distributed heterogeneous systems of the future with easy-to-program software is broadly acknowledged to be a grand challenge. It is widely recognized that a major disruption is currently under way in the design of computer systems as processors strive to extend, and go beyond, the end-game of Moore’s Law. This disruption is manifest in new forms of heterogeneous and distributed processors and memories at all scales (on-chip, on-die, on-node, on-rack, on-cluster, and on-data-center), rendering scalability as a fundamental challenge at all levels. Healthcare analytics offers a unique opportunity to explore scalable system design for the 21st century because there has been a tectonic shift in the ability of medical institutions to capture and store medical data, and to even stream data in real time. This shift has already contributed to an ecosystem of Machine Learning (ML) models being trained for a variety of clinical tasks. A new distributed heterogeneous architecture is required to build systems that can develop and deploy ML models based on distributed healthcare data that must necessarily be accessed with privacy-preserving constraints. Further, the proposed architecture must be accompanied by a software framework that can address the needs of domain-specific data scientists to develop and augment ML models being deployed in their hospitals.This planning grant project is exploring the foundational principles necessary in building integrated scalable distributed systems of the future, so as to prepare for submitting a full proposal to the PPoSS program. It uses the domain of healthcare analytics to motivate and concretize the research agenda, but the principles developed in this research should be applicable to other application domains as well. The exploration focuses on demonstrating an integrated platform that spans multiple levels of distribution and heterogeneity of computation and storage, while also obeying important privacy constraints. While recent progress on the use of ML in healthcare applications has been encouraging, current approaches do not a) scale to the degrees of parallelism, heterogeneity, and distribution that will be required in future systems, or b) support the soft real-time responsiveness to streaming data that is needed in many clinical situations. The originality of this project can be seen in the integration of distribution, heterogeneity, and privacy considerations in a single unified software/hardware stack, which includes adaptive resource management that spans privacy-preserving federated continuous learning, automatic specialization of ML models at individual sites, and automatic selection of ML models best suited for specific clinical tasks that maximize accuracy subject to different latency and soft real-time constraints.This project’s end-to-end approach to develop foundational scalability principles will impact multiple areas of computer science through publications, tutorials and courses, thereby benefiting other researchers working on scalability challenges in future distributed heterogeneous systems. The use of healthcare analytics as a driving application has the potential to result in significant benefits to society, by demonstrating how knowledge distilled from multiple sources of data can be embodied in recommendation systems that can run onsite to provide time-critical decision support to physicians. As a further impact, the project will contribute to the training of Highly Qualified Personnel (HQP) at the intersection of Systems for ML and ML for Healthcare — two emerging inter-disciplinary communities that are currently growing independent of each other. Finally, this research will leverage existing activities at the PIs’ institutions that contribute to broadening participation of underrepresented groups in computing.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.

人们普遍认为，使用易于编程的软件构建未来的可扩展分布式异构系统是一项巨大的挑战，随着处理器努力扩展和运行，计算机系统的设计目前正在发生重大变革。超越摩尔定律的最终阶段，这种破坏体现在各种规模的新形式的异构和分布式处理器和存储器中（片上、芯片上、节点上、机架上、集群上和存储器上）。数据中心上），将可扩展性视为各个层面的基本挑战，医疗保健分析为探索 21 世纪的可扩展系统设计提供了独特的机会，因为医疗机构捕获和存储医疗数据的能力发生了结构性转变。这种转变已经促成了针对各种临床任务进行训练的机器学习（ML）模型生态系统，以构建可以开发和部署的系统。基于分布式的ML模型此外，所提出的架构必须附带一个软件框架，该框架可以满足特定领域的数据科学家开发和增强在医院中部署的机器学习模型的需求。资助项目正在探索构建未来集成可扩展分布式系统的基本必要原则，以便为向 PPoSS 计划提交完整提案做好准备。它使用医疗保健分析领域来激励和具体化研究议程，但原则。本研究中开发的应该适用探索的重点是展示一个跨越计算和存储的多个分布和异构性的集成平台，同时遵守重要的隐私限制，尽管最近在医疗保健应用中使用机器学习的进展令人鼓舞。当前的方法不能a）扩展到未来系统所需的并行性、异质性和分布程度，或b）支持许多临床情况所需的对流数据的软实时响应。项目可以在集成中看到在单个统一的软件/硬件堆栈中考虑分布、异构性和隐私考虑因素，其中包括跨隐私保护联合持续学习的自适应资源管理、各个站点的 ML 模型的自动专业化，以及自动选择最适合特定情况的 ML 模型在不同的延迟和软实时约束下最大限度地提高准确性的临床任务。该项目开发基本可扩展性原则的端到端方法将通过出版物、教程和课程影响计算机科学的多个领域，从而使其他研究人员受益通过展示从多个数据源提取的知识如何体现在可以现场运行的推荐系统中，医疗保健分析作为驱动应用程序有可能为社会带来重大利益。作为进一步的影响，该项目将有助于在 ML 系统和医疗保健 ML 系统（两个目前正在独立发展的新兴跨学科社区）的交叉点培训高素质人员 (HQP)。最后，彼此。这项研究将利用 PI 机构的现有活动，有助于扩大代表性不足的群体在计算领域的参与。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

Enabling Real-time DNN Switching via Weight-Sharing

通过权重共享启用实时 DNN 切换

• 发表时间: 2022-07
• 期刊: Conference proceedings International Symposium on Computer Architecture
• 作者: Tong, Jianming;Chen, Yangyu;Pan, Yue;Bambhaniya, Abhimanyu;Khare, Alind;Heo, Taekyung;Tumanov, Alexey;Krishna, Tushar
• 通讯作者: Krishna, Tushar

SHMEM-ML: Leveraging OpenSHMEM and Apache Arrow for Scalable, Composable Machine Learning

SHMEM-ML：利用 OpenSHMEM 和 Apache Arrow 实现可扩展、可组合的机器学习

• 发表时间: 2022-05
• 期刊: Lecture notes in computer science
• 作者: Grossman, Ma;Poole, Steve;Pritchard, Howard;Sarkar, Vivek
• 通讯作者: Sarkar, Vivek