MRI: Acquisition of FlashTAIL - An All-NVMe Flash Storage Instrument for the Talon Artificial Intelligence & Machine Learning Cloud

该重大研究仪器(MRI)奖支持收购名为FlashTAIL的数据存储仪器，该仪器将由北达科他州大学(UND)管理。该仪器将为使用图形处理单元(GPU)计算加速器的高性能计算环境实现高速数据传输，并专门用于高级人工智能(AI)和机器学习(ML)应用。UD承诺在未来五年内进行重大投资，招聘新的教师，并形成一群在大数据、人工智能和ML相关领域具有专业知识的计算研究人员。FlashTAIL将允许这些研究人员和UND的其他教员在广泛的大学院系中发展数据科学、人工智能和ML研究能力。参与的研究人员提供的课程将吸引来自与集群教员有合作关系的系的学生。课程培训将融入由FlashTAIL仪器增强的GPU计算资源，为学生提供在尖端AI计算生态系统中工作的机会。通过这些途径，FlashTAIL将提高北达科他州研究的竞争力，为出现多样化的科技支持的地区劳动力做出贡献，并协助大学教职员工努力解决关于无人机系统(UAS)、大数据和人工智能、农村健康和能源可持续发展的大挑战目标。ML工作流程通常需要使用大型数据集来训练预测算法，作为强大的人工智能的基础。如果没有实现有效的数据流水线，则算法训练过程可能从计算受限(计算受限)问题转变为较慢的输入/输出(I/O受限)问题。这种性能状态称为“数据匮乏”，当处理器必须等待数据管道传递更多数据才能继续ML训练时，就会出现这种状态。训练过程越慢，可以完成的整体计算工作就越少，使得高级人工智能解决方案更难实施。收购FlashTAIL将有助于缓解“数据匮乏”对ML工作流程的影响，方法是为UND的OpenStack云系统Talon配备DataDirect Networks AI400存储仪器，该仪器能够以40 GB/S-20 GB/S的并发高输入/输出操作/秒数据速率传输大型数据集到Talon的两个支持Apollo 6500 NVLink GPU的计算节点(每个节点8个NVIDIA Tesla smx2 V100 GPU卡)。这将允许UND支持需要大型数据集来适当训练预测性ML算法的教师研究。这将提高UND研究开发计算机视觉人工智能应用程序的能力，这些应用程序用于为野生动物调查和建筑和基础设施分析等项目收集的大规模、大数据集UAS数据，以及用于评估用于生物组织开发的3D生物界面的小规模、大型数据集显微镜收集的数据。该奖项由NSF高级数字基础设施办公室(OAC)内的MRI和数据计划以及电气、光学和磁性设备司内的电子、光子和磁性设备(EPMD)计划管理并共同资助。这一奖项反映了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。

This Major Research Instrumentation (MRI) award supports the acquisition of a data storage instrument named FlashTAIL that will be managed by University of North Dakota (UND). This instrument will enable high speed data transfers for high-performance computing environments that use Graphical Processing Unit (GPU) computational accelerators and specialize in advanced Artificial Intelligence (AI) and Machine Learning (ML) applications. UND has committed to significant investments over the next five years to hire new faculty and form a cluster of computational researchers with specializations in areas relevant to Big Data, AI, and ML. FlashTAIL will allow these researchers and other faculty at UND to grow Data Science, AI, and ML research capabilities across a wide swath of university departments. Courses offered by the participating researchers will attract students from departments to which cluster faculty have collaborative relationships. Course training will incorporate the GPU computing resources enhanced by the FlashTAIL instrument, providing students with the opportunity to work within a cutting-edge AI computing ecosystem. Through these avenues, FlashTAIL will improve the competitiveness of North Dakota research, contribute to the emergence of a diverse tech-enabled regional workforce, and assist university faculty in their efforts to address the Grand Challenge objectives regarding unmanned aircraft systems (UAS), big data and AI, rural health, and energy sustainability.ML workflows often require the use of large datasets to train predictive algorithms that serve as the basis for robust AI. If an efficient data pipeline is not implemented, the algorithm training process can turn from a computationally restricted (compute bound) problem into a slower input/output (I/O bound) problem. This state of performance is known as "data starvation" and occurs when processors must wait for the data pipeline to deliver more data before the ML training can continue. The slower the training process, the less overall computational work can be completed, making advanced AI solutions more difficult to implement. Acquisition of FlashTAIL will serve to mitigate the impact of "data starvation" on ML workflows by equipping UND's OpenStack cloud system, Talon, with the DataDirect Networks AI400 storage instrument capable of streaming large datasets at a concurrent high input/output operations per second (IOPS) data rate of 40GB/s - 20GB/s to each of Talon's two HPE Apollo 6500 NVLink GPU-enabled compute nodes (8 x Nvidia Tesla smx2 V100 GPU cards per node). This will allow UND to support faculty research that requires large datasets to properly train predictive ML algorithms. This will improve the ability of UND research to develop computer vision AI applications for large-scale, large-dataset UAS-collected data for projects such as wildlife surveys and building and infrastructure analysis, and as well as small-scale, large-dataset microscopy-collected data for evaluating 3D biointerfaces for biological tissue developmentThis award is managed by and jointly funded through the MRI and Data programs within the NSF Office of Advanced Cyberinfrastructure (OAC) and the Electronics, Photonics and Magnetic Devices (EPMD) program within the Division of Electrical, Communications and Cyber Systems (ECCS).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.