OAC Core: Small: Next-Generation Communication and I/O Middleware for HPC and Deep Learning with Smart NICs

OAC 核心：小型：使用智能 NIC 实现 HPC 和深度学习的下一代通信和 I/O 中间件

• 批准号: 2007991
• 负责人: Dhabaleswar Panda
• 金额: $ 50万
• 依托单位: Ohio State University
• 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: 2020-07-01 至 2024-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2007991&HistoricalAwards=false
• 关键词: OAC; Core; Small; Next; Generation

In-network computing technologies, or the ability to offload significant portions of compute, communication, and I/O tasks to the network, have emerged as fundamental requirements to achieve extreme scale performance for end applications in the areas of High-Performance Computing (HPC) and Deep Learning (DL). Unfortunately, current generation communication middleware and applications cannot fully take advantage of these advances due to the lack of appropriate designs in the middleware-level. This leads to the following broad challenges: 1) Can middleware that are “aware” of the computing capabilities of these emerging in-network computing technologies be designed in the most optimized manner possible for HPC and DL applications?, and 2) Can such a middleware be used to benefit end applications in HPC and DL to achieve better performance and portability? A synergistic and comprehensive research plan is proposed to address the above broad challenges with innovative solutions. The proposed framework will be made available to collaborators and the broader scientific community to understand the impact of the proposed innovations on next-generation HPC and DL middleware and applications. Several graduate and undergraduate students will be trained under this project as future scientists and engineers in HPC. The proposed work will enable curriculum advancements via research in pedagogy for key courses at The Ohio State University. Tutorials and workshops will be organized at various conferences to share the research results and experience with the community. The project is aligned with the National Strategic Computing Initiative (NSCI) to advance US leadership in HPC and the recent initiative of the US Government to maintain leadership in Artificial Intelligence (AI.)The proposed innovations include: 1) Designing scalable communication primitives (point-to-point and collectives) for using emerging switch and NIC based in-network computing features, 2) Exploiting in-network computing features to offload complex and user defined functions, 3) Designing high-performance I/O and storage subsystems using NVMe over Fabrics, 4) Designing enhanced in-network datatype processing schemes for MPI library, 5) Designing and optimizing in-network computing-based solutions for emerging cloud environment, and 6) Carrying out integrated development and evaluation of the proposed designs with a set of representative HPC and DL applications. The proposed designs will be integrated into the widely-used MVAPICH2 library and made available to the public. The project team members will work closely with collaborators to facilitate wide deployment and adoption of released software. The transformative impact of the proposed research is to achieve scalability, performance, and portability for HPC and DL frameworks/applications by leveraging emerging in-network computing technologies.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.

网络内计算技术，即将大部分计算、通信和I/O任务卸载到网络的能力，已成为高性能计算(HPC)和深度学习(DL)领域终端应用实现极高规模性能的基本要求。遗憾的是，由于在中间件级别缺乏适当的设计，当前一代的通信中间件和应用程序不能充分利用这些进步。这带来了以下广泛的挑战：1)“意识到”这些新兴的网络计算技术的计算能力的中间件能否以尽可能优化的方式设计用于高性能计算机和数字图书馆应用程序？2)这样的中间件能否用于优化高性能计算机和数字图书馆中的fi终端应用程序，以实现更好的性能和可移植性？提出了一项协同和全面的研究计划，以创新的解决方案应对上述广泛的挑战。建议的框架将向合作者和更广泛的科学界提供，以了解建议的创新对下一代HPC和DL中间件和应用程序的影响。几名研究生和本科生将在该项目下接受培训，成为HPC的未来科学家和工程师。这项拟议的工作将通过对俄亥俄州立大学关键课程的教育学研究来推进课程。将在不同的会议上组织教程和工作坊，与社会各界分享研究成果和经验。该项目与推进美国在高性能计算领域的领导地位的国家战略计算倡议和美国政府最近为保持在人工智能(AI)领域的领导地位的倡议保持一致。提出的创新包括：1)设计可扩展的通信原语(点对点和集体)，用于使用新兴的基于交换机和NIC的网络内计算功能；2)利用网络内计算特征来实现fflOAD Complex和User defiNed功能；3)使用NVMe over Fabric设计高性能的I/O和存储子系统；4)为MPI库设计增强的网络内数据类型处理方案，5)针对新兴的云环境设计和优化基于网内计算的解决方案；6)结合一组具有代表性的HPC和DL应用对所提出的设计进行集成开发和评估。建议的设计将被整合到广泛使用的MVAPICH2库中，并向公众提供。项目团队成员将与合作者密切合作，促进已发布软件的广泛部署和采用。拟议研究的变革性影响是通过利用新兴的网络内计算技术实现HPC和DL框架/应用程序的可扩展性、性能和可移植性。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

Hy-Fi: Hybrid Five-Dimensional Parallel DNN Training on High-Performance GPU Clusters

Hy-Fi：高性能 GPU 集群上的混合五维并行 DNN 训练

• DOI: 10.1007/978-3-031-07312-0_6
• 发表时间: 2022
• 期刊: Proceedings International Conference on High Performance Computing
• 作者: Jain, A;Shafi, A.;Anthony, Q.;Kousha, P.;Subramoni, H.;Panda, DK.
• 通讯作者: Panda, DK.

Highly Efficient Alltoall and Alltoallv Communication Algorithms for GPU Systems

适用于 GPU 系统的高效 Alltoall 和 Alltoallv 通信算法

• DOI: 10.1109/ipdpsw55747.2022.00014
• 发表时间: 2022
• 期刊: Heterogeneity in Computing Workshop
• 作者: Chen, Chen-Chun;Khorassani, Kawthar Shafie;Anthony, Quentin G.;Shafi, Aamir;Subramoni, Hari;Panda, Dhabaleswar K.
• 通讯作者: Panda, Dhabaleswar K.

Network-Assisted Noncontiguous Transfers for GPU-Aware MPI Libraries

GPU 感知 MPI 库的网络辅助非连续传输

• DOI: 10.1109/mm.2023.3241133
• 发表时间: 2023
• 期刊: IEEE Micro
• 影响因子: 3.6
• 作者: Suresh, Kaushik Kandadi;Khorassani, Kawthar Shafie;Chen, Chen Chun;Ramesh, Bharath;Abduljabbar, Mustafa;Shafi, Aamir;Subramoni, Hari;Panda, Dhabaleswar K.
• 通讯作者: Panda, Dhabaleswar K.

High Performance MPI over the Slingshot Interconnect: Early Experiences

基于 Slingshot 互连的高性能 MPI：早期经验

• DOI: 10.1145/3491418.3530773
• 发表时间: 2022
• 期刊: Practice and Experience in Advanced Research Computing
• 作者: Shafie Khorassani, Kawthar;Chen, Chen Chun;Ramesh, Bharath;Shafi, Aamir;Subramoni, Hari;Panda, Dhabaleswar
• 通讯作者: Panda, Dhabaleswar

OMB-Py: Python Micro-Benchmarks for Evaluating Performance of MPI Libraries on HPC Systems

OMB-Py：用于评估 HPC 系统上 MPI 库性能的 Python 微基准

• DOI: 10.1109/ipdpsw55747.2022.00143
• 发表时间: 2022
• 期刊: 23rd Parallel and Distributed Scientific and Engineering Computing Workshop (PDSEC
• 作者: Alnaasan, Nawras;Jain, Arpan;Shafi, Aamir;Subramoni, Hari;Panda, Dhabaleswar K
• 通讯作者: Panda, Dhabaleswar K