CAREER: Machine Learning Driven Cross-Layer Optimizations for Storage

职业：机器学习驱动的跨层存储优化

• 批准号: 1942754
• 负责人: Heiner Litz
• 金额: $ 53万
• 依托单位: University of California-Santa Cruz
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-06-15 至 2025-05-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1942754&HistoricalAwards=false
• 关键词: CAREER; Machine; Learning; Driven; Cross

Recent advancements in computer science have enabled an exponential data growth outpacing technology scaling. As an increasing number of mobile devices, sensors and data acquisition systems is producing exabytes of information, analyzing the obtained data are becoming unfeasible, requiring novel approaches to store and compute on this vast amount of data. Improving the performance and efficiency of storage systems is of paramount importance to enable scientific progress, to improve the cost and energy consumption of IT systems and to enable analysis of large amounts of data. To achieve this goal, as part of this grant, novel approaches on the hardware, operating systems, data-center and application level will be developed. Enabling such new techniques will provide significant benefit for society. First, the new approaches developed in this project will improve efficiency and utilization of storage systems reducing the carbon footprint on our world. Secondly, improving the performance of storage devices enables novel applications such as new treatments leveraging storage and compute intensive genomics.As part of this project, cross layer optimizations will be developed to improve the hardware and software stack of storage systems using machine learning techniques. One main challenge of existing block storage devices is their transparency of internal state to software. This work addresses this shortcoming by extending storage devices with comprehensive data monitoring capabilities as well as with control knobs to optimize devices in an application specific way. The telemetry data obtained from these smart storage devices will be utilized to train machine learning models to optimize for application specific behavior as well as for determining optimal configurations of heterogeneous storage and compute environments.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.

计算机科学的最新进展使得数据呈指数级增长，超过了技术扩展的速度。随着越来越多的移动设备、传感器和数据采集系统产生数亿字节的信息，分析所获得的数据变得不可行，需要新的方法来存储和计算如此大量的数据。提高存储系统的性能和效率对于推动科学进步、降低IT系统的成本和能耗以及实现大量数据的分析至关重要。为了实现这一目标，作为这笔拨款的一部分，将开发硬件、操作系统、数据中心和应用程序级别的新方法。启用此类新技术将为社会带来重大利益。首先，该项目开发的新方法将提高存储系统的效率和利用率，减少我们世界的碳足迹。其次，提高存储设备的性能可以实现新颖的应用，例如利用存储和计算密集型基因组学的新治疗方法。作为该项目的一部分，将开发跨层优化，以使用机器学习技术改进存储系统的硬件和软件堆栈。现有块存储设备的主要挑战之一是其内部状态对软件的透明度。这项工作通过扩展存储设备来解决这一缺点，提供全面的数据监控功能以及控制旋钮，以特定于应用的方式优化设备。从这些智能存储设备获得的遥测数据将用于训练机器学习模型，以优化应用程序的特定行为以及确定异构存储和计算环境的最佳配置。该奖项反映了 NSF 的法定使命，并被认为值得支持通过使用基金会的智力优点和更广泛的影响审查标准进行评估。

项目成果

RAIL: Predictable, Low Tail Latency for NVMe Flash

• DOI: 10.1145/3465406
• 发表时间: 2022-01
• 期刊: ACM Transactions on Storage (TOS)
• 作者: Heiner Litz;Javier González;Ana Klimovic;Christos Kozyrakis

Online Code Layout Optimizations via OCOLOS

通过 OCOLOS 在线代码布局优化

• DOI: 10.1109/mm.2023.3274758
• 发表时间: 2023
• 期刊: IEEE Micro
• 影响因子: 3.6
• 作者: Zhang, Yuxuan;Khan, Tanvir Ahmed;Pokam, Gilles;Kasikci, Baris;Litz, Heiner;Devietti, Joseph
• 通讯作者: Devietti, Joseph

Predict and Write: Using K-Means Clustering to Extend the Lifetime of NVM Storage

• DOI: 10.1109/icde51399.2021.00072
• 发表时间: 2021-01-01
• 期刊: 2021 IEEE 37TH INTERNATIONAL CONFERENCE ON DATA ENGINEERING (ICDE 2021)
• 作者: Kargar, Saeed;Litz, Heiner;Nawab, Faisal
• 通讯作者: Nawab, Faisal

Deep Learning based Prefetching for Flash

基于深度学习的 Flash 预取

• 发表时间: 2022
• 期刊: Nonvolatile Memory Workshop (NVMW
• 作者: Chakraborttii, Chandranil;Litz, Heiner
• 通讯作者: Litz, Heiner

APT-GET: profile-guided timely software prefetching

• DOI: 10.1145/3492321.3519583
• 发表时间: 2022-03
• 期刊: Proceedings of the Seventeenth European Conference on Computer Systems
• 作者: Saba Jamilan;Tanvir Ahmed Khan;Grant Ayers;Baris Kasikci;Heiner Litz