CRII: SHF: Optimizing Program Executions on Non-uniform Threaded Architectures

现代计算机系统采用多线程来利用具有多个功能单元的多个核。高线程级并行性源于每个内核具有多个线程（同时或细粒度多线程，SMT）、每个处理器具有多个内核（芯片多线程处理器，CMP）以及每个节点具有多个处理器（非均匀内存访问，NUMA）。线程根据它们执行的位置共享不同级别的硬件资源。具有混合SMT、CMP和NUMA线程的体系结构是非统一线程体系结构。今天的大多数多插槽系统都是这种形式。由于操作系统和编程模型中的线程抽象，软件开发人员通常将系统中的所有线程视为对称的，而忽略了它们在硬件中的非一致性。因此，在非统一线程架构上运行的多线程代码的性能远远低于理论峰值，这可能会降低生产率并增加能耗。这个项目的主要重点是研究静态和动态软件方法，以利用硬件线程之间的非一致性。对于静态优化，PI旨在通过代码转换在软件中引入线程非一致性。这还包括通过在应用程序源代码中确定应用此类转换的机会来改进现有应用程序。对于动态优化，PI计划研究在线调度方法，以匹配软件和硬件中的非均匀线程。给定一个可执行文件，它将被在线分析以表征其线程之间的资源共享，并将对线程和数据应用适当的调度策略。这个项目将紧密结合静态和动态优化方法的程序运行在非统一的线程架构。这项研究可以显着提高性能的大规模多线程应用程序运行在今天？和新兴的并行架构。更广泛地说，这个项目将在设计性能工具和非统一线程的并行编程框架方面产生重大影响。它可能会引起工业界和学术界的广泛兴趣。该项目的一个重要组成部分是它与本科和研究生课程教学以及学生辅导的整合。

Modern computer systems employ multiple threads to exploit multiple cores with multiple functional units. High thread-level parallelism arises from having multiple threads per core (simultaneous or fine-grained multithreading, SMT), multiple cores per processor (chip multithreaded processor, CMP), and multiple processors per node (non-uniform memory access, NUMA). Threads share different levels of hardware resources depending on where they execute. An architecture with hybrid SMT, CMP and NUMA threads is a non-uniform threaded architecture. Most multi-socket systems today are of this form. Due to the thread abstraction in operating systems and programming models, software developers often treat all threads in the system as symmetric, ignoring their non-uniformity in hardware. As a result, multithreaded code running on non-uniform threaded architectures performs at a level far below the theoretical peak, which can degrade productivity and increase energy consumption. The main focus of this project is to investigate both static and dynamic software methods to exploit non-uniformity between hardware threads. For the static optimization, the PI aims to introduce thread non-uniformity in software via code transformations. This will also include improving existing applications by identifying opportunities in application source code to apply such transformations. For the dynamic optimization, the PI plans to study online scheduling methods to match non-uniform threads in software and hardware. Given an executable, it will be analyzed online to characterize resource sharing between its threads and appropriate scheduling strategies will be applied for both threads and data. This project will tightly integrate static and dynamic optimization methods for programs running on non-uniform threaded architectures. This research can dramatically improve the performance of large-scale multithreaded applications running on today?s and emerging parallel architectures. More broadly, this project will have a strong impact in designing performance tools and parallel programming frameworks for non-uniform threads. It will likely attract broad interest from industry and academia. An important part of this project is its integration with teaching undergraduate and graduate courses as well as student mentoring.