CAREER: Application-Driven Memory Hierarchy Management

CCR-9876128 CAREER：应用程序驱动的内存层次结构管理Vijay Karamcheti处理器和内存性能之间不断扩大的差距降低了需求驱动缓存等反应式内存层次结构管理技术的有效性。应用程序驱动的主动技术(如预取)已被证明是有效的，但需要程序员的广泛参与。该提案研究了在一般编程环境中实现此类技术的自动化，特别是在商用工作站集群上的基于对象的分布式共享内存(DSM)实现。集成的编译器和运行时技术首先建立程序中线程交互的模型，然后使用该模型构建优化DSM性能的每线程自定义一致性协议，并结合运行时分析确定当前活动的线程部分及其副作用。每线程定制协议依靠灵活的运行时一致性和调度原语、推测性优化以及对可用的结构支持的改进利用，克服了编译时的不精确性。教育计划侧重于强调现有计算机系统课程中应用程序、系统软件和硬件之间的相互作用，并开发新的自适应计算机系统实践课程(允许在编译器和/或应用程序控制下进行硬件和软件的配置)。预期的项目成果包括自动应用程序驱动的内存层次管理技术，将其整合到编译器和运行时基础设施中以优化集群平台上的程序性能，以及接受应用程序感知计算机系统设计培训的学生。

CCR-9876128CAREER: Application-Driven Memory Hierarchy ManagementVijay KaramchetiThe widening gap between processor and memory performance has reducedthe effectiveness of reactive memory hierarchy management techniquessuch as demand-driven caching. Application-driven proactivetechniques (such as prefetching) have been shown to be effective, butrequire extensive programmer involvement. This proposal investigatesautomating such techniques for a general programming context,specifically object-based distributed shared-memory (DSM)implementations on clusters of commodity workstations. Integratedcompiler and run-time techniques first develop a model of threadinteractions in the program, and then, using this model, constructper-thread custom coherence protocols that optimize DSM performance.Static analyses augmented with run-time profiling determine theconcurrently-active thread portions and their side-effects. Theper-thread custom protocols overcome compile-time imprecision byrelying on flexible run-time coherence and scheduling primitives,speculative optimizations, and improved utilization of availablearchitectural support. Educational plans focus on emphasizing theinterplay among applications, system software, and hardware inexisting computer systems courses, and developing new hands-on coursesin adaptive computer systems (which allow hardware and softwareconfigurability under compiler and/or application control). Expectedproject outcomes include technology for automatic application-drivenmemory-hierarchy management, its incorporation into compiler andrun-time infrastructure for optimizing program performance on clusterplatforms, and students trained in application-aware computer systemdesign.