SHF: Small: Revamping I/O Architectures Using Machine Learning Techniques on Big Compute Machines

在现代计算机系统中，快速增长的计算速度和缓慢改善的片外存储器和磁盘驱动器之间的数据传输速率之间的差距是一个长期存在的研究挑战，通常被称为输入/输出墙问题。在如今的大数据、大计算时代，这个问题已经变得十分严峻。尽管近年来数据存储技术发展迅速，但机器和工作负载的异质性和多样性日益增加，再加上持续的数据爆炸，加剧了计算和磁盘驱动器存储之间的速度差距。人们越来越需要为新兴的大规模和多样化的应用开发一种高性能、经济高效的体系结构，而不受输入/输出墙问题的影响。该项目的目标是利用现有的大型计算资源，如图形处理单元（gpu）和深度学习技术，在不增加新硬件的情况下加快二级存储系统的性能。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。本项目建议使用人工智能开发新的支持gpu的在线学习I/O架构。该项目包括三个研究重点：首先，它将为存储系统开发定制的机器学习和深度学习算法和模型。例如，它将设计一种时间感知分类技术来解决极端尺度学习问题。二是针对预取、日志管理、垃圾回收等核心存储系统模块开发新的学习解决方案。第三，它将这些建议的交织模块集成到一个异构GPU机器和GPU集群中。它将为内部固态磁盘（SSD）设备和非易失性内存快速（NVMe）和外围连接接口快速（PCIe）协议开发最佳并行解决方案，以构建快速通道，在存储和gpu之间直接移动数据。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。

In modern computer systems, the growing disparity in rapidly increasing computational speeds and slowly improving data transfer rates to/from off-chip memory and disk drives is a long-standing research challenge, often referred to as the Input/Output wall problem. This problem has become severe in today's big data and big compute era. Despite the rapid evolution in data storage technologies in recent years, the increasing heterogeneity and diversity in machines and workloads, coupled with the continued data explosion, exacerbate the speed gap between computing and disk-drive storage. There is an increasing need to develop a high-performance, and cost-effective architecture for emerging large-scale and diverse applications that is not affected by the Input/Output wall problem. The goal of this project is to leverage existing big compute resources such as graphic processing units (GPUs) and deep learning techniques to speed up the secondary storage system performance without adding new hardware. 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. This project proposes to develop new GPU-enabled online learned I/O architecture using artificial intelligence. This project entails three research thrusts: First, it will develop custom machine learning and deep learning algorithms and models for the storage system. For example, it will design a temporal-aware classification technique to attack the extreme-scale learning problem. Second, it will develop new learning solutions for core storage system modules such as prefetching, log management and garbage collection. Third, it will integrate these proposed interwoven modules into a heterogeneous GPU machine and a GPU cluster at scale. It will develop optimal parallelism solutions for internal solid-state disk (SSD) devices and Non-Volatile Memory Express (NVMe) and Peripheral Connection Interface Express (PCIe) protocol to construct an express channel, moving data directly between storage and GPUs.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.