CSR: Small: Collaborative Research: Tuning Extreme-scale Storage Stack through Deep Reinforcement Learning

许多研究领域，如高能物理、气候科学、天体物理学、燃烧科学和计算生物学，都需要处理大量的数据。这些领域严重依赖高性能计算（HPC）系统的能力来管理和有效处理大量数据。因此，上述研究领域的应用需要高性能计算存储系统的高度优化性能，用于存储、管理和操作数据。本项目旨在利用深度强化学习方法对高性能计算存储系统进行微调，以优化性能。本研究探讨了利用深度强化学习优化HPC存储系统的可行性：(a)创建基于深度学习的HPC存储堆栈模型；(b)改造现有的高性能计算存储栈，以支持自动化配置和调优；(c)收集训练数据集和训练存储堆栈模型；(d)利用模型作为响应性和可玩性的虚拟环境来学习调整参数的最佳策略。作为一个合作项目，本研究旨在推进HPC存储系统和机器学习的领域知识。高性能计算存储堆栈的性能增强将反过来有利于科学发现，从而有利于我们的社会。研究人员将在项目过程中整合研究、教育和推广工作，包括招募和留住代表性不足的学生，指导研究生和本科生，将研究成果整合到课程中，并出版和传播结果。用于训练存储栈模型的数据将在https://discl.cs.ttu.edu/tuningstorage上共享，而机器学习的代码将在https://github.com/forrestbao/DL4SC上共享。结果和数据将在出版时提供。数据将作适当注释，以方便解释。主要研究人员将努力尽可能长时间地维护存储库。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Many research domains, such as high-energy physics, climate science, astrophysics, combustion science, and computational biology, need to process large amounts of data. Such domains are heavily relying on the capabilities of high performance computing (HPC) systems to manage and efficiently process massive amounts of data. Consequently, applications in the aforementioned research domains require highly optimized performance on the HPC storage systems that store, manage, and manipulate data. This project aims to utilize deep reinforcement learning methods to fine-tune the HPC storage system for optimized performance.This research explores the feasibility of leveraging deep reinforcement learning to optimize HPC storage systems by: (a) Creating a deep learning based HPC storage stack model; (b) Remodeling existing HPC storage stack to support automated configuration and tuning; (c) Collecting training datasets and training the storage stack model; and (d) utilizing the model as a responsive and playable virtual environment to learn the best policy to tune parameters. As a collaborative project, this research aims to advance the domain knowledge of both HPC storage systems and machine learning. The enhanced performance on the HPC storage stack will in turn benefit scientific discovery and thus our society. The investigators will integrate research, education, and outreach efforts during the course of this project, including recruiting and retaining of underrepresented students, mentoring graduate and undergraduate students, integrating research findings into curriculum, and publishing and disseminating results.The data collected to train the storage stack model will be shared at https://discl.cs.ttu.edu/tuningstorage while the code of machine learning at https://github.com/forrestbao/DL4SC. Results and data will be made available by the time of publication. The data will be annotated as appropriate to facilitate interpretation. The principal investigators will strive to maintain the repositories as long as possible.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.