CAPA: Collaborative Research: Lightweight Abstract Memory Features

现代计算系统的内存子系统已经经历了巨大的创新，它结合了新的特性来帮助程序员创建快速、安全、正确和高效的软件。不幸的是，利用这些特性是一个挑战，因为编程语言没有向程序员公开内存子系统的高级功能。这个项目的重点是编程语言和存储硬件之间的接口。智力上的优点包括为内存特性创建严格的语义，这样程序员就可以对单个特性及其组成进行推理；以及编译工具和运行时系统的创建，这些工具和运行时系统允许程序员单独或组合使用新的内存特性，以获得最大的收益。该项目更广泛的意义和重要性在于它对软件和硬件行业的影响，促进了新内存功能的快速采用；研究生的教育与培养；并创建教程，以帮助传播和采用已开发的工具和技术。内存特性的硬件实现通常很快，但受到物理容量的限制，并且特性的精确语义通常是特定于供应商的。这项工作的理论部分将为内存特性创建严格的语义，它超越了个体实现，并允许程序员和静态分析工具推断程序与内存的交互。该研究将构建虚拟运行时系统来克服硬件限制，并在不存在的情况下模拟功能。它的实现将使用运行时适应性来微调自己，以适应给定系统的功能可用性、对功能组合的支持以及硬件/工作负载特征。构建在LLVM系统之上的自定义编译器基础结构将提供轻量级语法，程序员可以通过该语法轻松地在现有代码中添加对内存特性的支持。

The memory subsystem of modern computing systems has seen tremendous innovations, incorporating new features to aid programmers in creating fast, secure, correct, and power-efficient software. Unfortunately, harnessing these features is a challenge, as programming languages do not expose advanced abilities of the memory subsystem to programmers. This project focuses on the interface between programming languages and memory hardware. The intellectual merits include the creation of a rigorous semantics for memory features, so that programmers can reason about individual features and their composition; and the creation of compilation tools and run-time systems that allow programmers to use new memory features, in isolation or combination, for maximum gain. The project's broader significance and importance are its impact on the software and hardware industry, by facilitating rapid adoption of new memory features; the education and training of graduate students; and the creation of a tutorial to aid in dissemination and adoption of the developed tools and techniques.Hardware implementations of memory features are typically fast but limited by physical capacity, and the precise semantics of features are often vendor-specific. The theoretical portion of the work will create rigorous semantics for memory features, which transcend individual implementation and allow programmers and static analysis tools to reason about a program's interaction with memory. The research will construct virtualized run-time systems to overcome hardware constraints, and to emulate features when they are not present. Its implementations will employ run-time adaptivity to fine-tune themselves to a given system's feature availability, support for composition of features, and hardware/workload characteristics. A custom compiler infrastructure, built atop the LLVM system, will provide a lightweight syntax through which programmers can easily add support for memory features to their existing codes.