SPX: Collaborative Research: Memory Fabric: Data Management for Large-scale Hybrid Memory Systems

正在为国家实验室和美国工业界开发的新的大规模高性能计算系统，结合了不同的存储器组件、加速器和加速器-近存储器以及可编程的高性能互连。这些内存丰富的设计很有吸引力，因为它们提供了缩短科学发现时间和支持新的延迟敏感型数据密集型应用程序所需的接近数据的计算容量。然而，现有的软件堆栈没有配备来处理这些机器设计的异构性和复杂性，这会影响应用程序性能和机器效率。在本项目中开发的内存结构(MF)解决方案提供了新的抽象和机制，允许系统软件堆栈更深入地了解应用程序的数据使用模式和要求，并协调有关如何在不同的内存中分布数据或沿不同的互连路径交换数据的决策。内存结构(MF)体系结构引入了新的以数据为中心的抽象、内存对象和内存对象流，以及伴随而来的内存和通信管理方法。在新的抽象中捕获的更高级别的信息使MF运行时能够更好地指导底层内存和互连管理，并掩盖底层内存底层的复杂性。使用近内存结构计算还有其他好处，包括通过动态插入特定于应用程序的代码，从而进一步专门化和加速由MF执行的操作。MF使用几个重要的应用领域进行评估，包括大数据学习和分析以及传统的高性能科学模拟。它的好处包括应用程序性能和资源效率的提高，同时保护应用程序和应用程序开发人员不受底层机器细节的影响。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

New large-scale high performance computing systems being developed for the national labs and by US industry, combine heterogeneous memory components, accelerators and accelerator-near memory, and programmable high-performance interconnects. These memory-rich designs are attractive as they provide the compute-near-data capacity needed for improving the time to scientific discovery, and for supporting new classes of latency-sensitive data-intensive applications. However, existing software stacks are not equipped to deal with the heterogeneity and complexity of these machine designs, which impacts application performance and machine efficiency. The Memory Fabric (MF) solution developed in this project provides new abstractions and mechanisms that permit the systems software stacks to gain deeper insight into applications' data usage patterns and requirements, and to coordinate the decisions concerning how data should be distributed across different memories, or exchanged along different interconnection paths. The Memory Fabric (MF) architecture introduces new data-centric abstractions, memory object and memory object flow, and accompanying memory and communications management methods. The higher-level information captured in the new abstractions empowers the MF runtime to better guide the underlying memory and interconnect management, and to mask the complexities of the underlying memory substrate. Additional benefits are derived from use of near-memory-fabric computation, including via dynamically inserted application-specific codes, which further specialize and accelerate the operations carried out by MF. MF is evaluated using several important application domains, including big data learning and analytics, and traditional high-performance scientific simulations. Its benefits include gains in application performance and resource efficiency, while shielding applications and application developers from the underlying machine details.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.