PPoSS: Planning: Data Centric Computing for Scalable Heterogeneous Memory and Storage Systems Architecture

该奖项支持一个大型项目的规划，解决由许多并行内存库和磁盘组成的异构内存和存储系统之间的可扩展性问题，特别是大数据和深度学习应用。随着时间的推移，新兴的大数据和学习应用显著增加了计算机系统中数据传输的频率和数量，从而导致大规模并行内存存储系统中不可忽视的争用和负载不平衡问题。这样的担忧可能会破坏计算的可伸缩性。为了达到对现实世界产生影响的必要规模，需要一种跨层方法来优化整个硬件/软件堆栈。该项目支持跨学科的规划过程，研究人员来自计算机体系结构、高性能计算（HPC）、安全和隐私、系统和算法。该项目还将通过吸纳西班牙裔服务机构中代表性不足的群体，以及传播计算机科学和工程教育与培训方面的研究成果，为社会做出贡献。该基金将计划建立一个研究团队，有能力解决大规模异构内存和存储系统的跨学科可扩展性挑战。pi利用应用程序提示，例如在完整的硬件/软件堆栈上额外的采样灵活性，以减少内存和输入/输出之间的数据传输流量。这是为了实现最终目标，即在低级内存和存储中维护无共享架构，从而实现高可伸缩性。具体的研究重点是：1)为自动驾驶和高性能计算等各种应用开展工作负载和应用特性研究；2)探索一种创新的方式，为并行内存库和磁盘开发一种应用感知的、以数据为中心的任务调度器，以实现高可扩展性；3)开发几种机器学习算法，以增强任务调度；4)研究了敏感数据集的同态加密在数据分析和采样中的应用。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。

This award supports planning for a large-grant project addressing the scalability issue among heterogeneous memory and storage systems that are composed of many parallel memory banks and disks, specifically for big-data and deep-learning applications. Emerging big-data and learning applications significantly increase both frequency and volume of data transfers over time in computer systems, which consequently result in non-negligible contention and load-imbalance issues among large-scale parallel memory storage systems. Such concerns can derail computing scalability. To grow to the requisite scale for real-world impact, there is a need for a cross-layer approach that optimizes the entire hardware/software stack. This project supports an interdisciplinary planning process, with researchers from computer architecture, high-performance computing (HPC), security and privacy, systems, and algorithms. This project will also contribute to society through engaging under-represented groups from a Hispanic Serving Institution and research dissemination for computer science and engineering education and training.The grant will plan to build a research team with the capability to address the interdisciplinary scalability challenges to large-scale heterogeneous memory and storage systems. The PIs exploit application hints such as additional sampling flexibility at the full hardware/software stack to reduce the amount of data-transfer traffic among memory and Input/Output. This is intended to realize the ultimate objective, namely maintaining a shared-nothing architecture in low-level memory and storage and thus achieving high scalability. Specific research thrusts are: 1) conducting workload and application-characteristics studies for various applications such as autonomous driving and high-performance computing, 2) exploring an innovative way to develop an application-aware, data-centric task scheduler for parallel memory banks and disks to achieve high-scalability, 3) developing several machine-learning algorithms to enhance task scheduling, and 4) investigating the applications of homomorphic encryption for sensitive datasets in data analytics and sampling applications.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.