SHF: Small: Automatic, adaptive and massive parallel data processing on GPU/RDMA clusters in both synchronous and asynchronous modes

硬件和软件的计算生态系统正处于一个关键的过渡时期，来自几个技术危机和不可避免的趋势。首先，通用处理器的持续性能改进不再现实。其次，对于各种数据密集型应用程序，传统处理器在性能和功耗方面的效率越来越低。最后，已经发展了几十年的深度软件堆栈，从指令集架构一直到现有生态系统中的编程层，增加了繁琐的处理，甚至在计算中增加了不必要的开销。为了解决上述问题，该项目以基于加速器的方式修复了计算生态系统。GPU（图形处理单元）和RDMA（远程数据存储器访问）是项目中考虑的两个外部硬件加速器。它旨在通过消除现有生态系统中的三个技术障碍(1)编程模型障碍，(2)硬件抽象障碍，(3)自动化障碍，自适应和自动地将高效的异步计算变成GPU和RDMA硬件加速器集群。该项目力求产生广泛和变革性的影响。预计它将以新的算法和有效的系统实施影响数据处理研究界，并影响行业在日常计算操作中改善其生产系统，为社会服务。开发的算法、源代码和测量方法可在网上广泛使用，使工业和学术研究人员受益。对本科生和研究生的研究训练解决了信息技术和计算行业缺乏硬件加速和数据分析专业人员的问题。课程开发将相关研究成果引入课堂，外展活动鼓励高中生对计算机相关大学教育感兴趣。现有的计算环境不提供异步执行的编程模型。在GPU/RDMA集群上进行异步编程更加困难。由于cpu和GPU的执行模式不同，使得系统缺乏一个通用的硬件抽象来进行GPU计算和RDMA通信和管理。异步编程很困难，因此非常需要一个确保其正确性和效率的自动工具。这个研究项目弥合了异步计算和GPU/RDMA之间的差距。开发了一个独立内存池（AMP）接口GPU/RDMA集群，其中提出了一个中间表示抽象GPU执行和由RDMA构造的AMP。开发了一套支持异步编程的中间表示，使用户可以方便地在编程中表示异步计算。此外，还开发了一种中间表示，允许传统的同步编程变成自动异步执行代码。该系统在大型GPU/RDMA集群上使用代表性的数据处理工作负载进行测试和评估。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。

The computing ecosystem in both hardware and software is in a critical transition time, coming from several technology crises and inevitable trends. First, the continued performance improvement in general-purpose processors is no longer realistic. Second, conventional processors are increasingly inefficient in both performance and power consumption for various data-intensive applications. Finally, the deep software stack that has been developed for several decades, from instruction-set architecture all the way to the programming layer in the existing ecosystem, has added cumbersome processing and even unnecessary overhead in computing. To address the above-mentioned issues, this project remedies the computing ecosystem in an accelerator-based way. GPU (Graphic Processing Unit) and RDMA (Remote Data Memory Access) are the two external hardware accelerators considered in the project. It aims to turn efficient asynchronous computing into a reality on clusters of hardware accelerators of GPU and RDMA adaptively and automatically by removing three technical barriers in the existing ecosystem: (1) the programming-model barrier, (2) the hardware abstraction barrier, and (3) the automation barrier. The project strives to make broad and transformational impact. It is expected to influence the data-processing research community with new algorithms and effective systems implementation, and influence industries to improve their production systems in their daily computing operations serving society. The developed algorithms, source code and measurements are available online for a public and wide usage, benefiting both industrial and academic researchers. The research training to both undergraduate and graduate students address the concerns of lacking hardware-acceleration and data-analytics professionals in information technology and computing industries. The curriculum development introduces related research results to classrooms and the outreach activities encourage high school students to be interested in computing related college education. The existing computing environment does not provide programming models for asynchronous execution. It is even harder for asynchronous programming on GPU/RDMA clusters. The execution-model difference between CPUs and GPUs makes the system lack a common hardware abstraction for GPU computing and for RDMA communication and management. Asynchronous programming is hard, and an automatic tool to ensure its correctness and efficiency is highly desirable. This research project bridges the gap between asynchronous computing and GPU/RDMA. It develops an autonomous memory pool (AMP) interfacing GPU/RDMA clusters, where an intermediate representation is proposed to abstract the GPU execution and AMP constructed by an RDMA. A set of intermediate representations are developed to support asynchronous programming, so that users can easily express asynchronous computing in programming. In addition, an intermediate representation is developed to allow conventional synchronous programming to become automated asynchronous execution code. The system is tested and evaluated using representative data-processing workloads on large GPU/RDMA clusters.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.