Integrated application and operating system optimization

该研究项目的主要目标是通过动态优化跨多个软件层的执行路径，使应用程序能够实现新兴计算机硬件系统的性能潜力。实现这一目标的主要挑战来自于计算机系统日益增长的复杂性，以及为帮助管理这种复杂性而开发的许多抽象层。两种技术趋势凸显了这个问题。首先，新兴的微处理器（例如Sun的多核多线程Niagara）具有非常难以理解和利用的微妙性能特征，即使是专家级程序员也很难理解和利用。与此同时，程序员正在使用诸如Java之类的语言，其中代码由抽象机器解释。这些语言提供的生产力提高部分是基于更高层次的抽象，它将程序员从实际硬件的令人不快的细节中解放出来。在程序员编写的代码和实际硬件执行的指令之间存在许多其他层，包括语言级虚拟机、操作系统、低级虚拟机和虚拟机监视器。这种情况很可能出现在当前的网络托管中心。我们的研究探索了在运行时打破静态抽象层所带来的障碍的各种方法。作为一个特定的项目，我们将探索将基于编译器的内存管理算法集成到Java JIT编译器中，以便生成的代码可以与JVM的堆管理系统交互，这将反过来通知操作系统页面预取和替换的候选对象。作为第二个特定的项目，我们将研究在操作系统代码本身中使用动态生成的执行跟踪（过去的项目仅将这些技术应用于用户级代码）。我们最初的目标是支持增强的概要分析和操作系统的性能检测，其次是动态优化操作系统中的执行路径，最终目标是集成用户级和系统级的执行跟踪。

The primary objective of this research project is to allow applications to realize the performance potential of emerging computer hardware systems, by dynamically optimizing execution paths across multiple software layers. The main challenge in achieving this objective arises from the growing complexity of computer systems and the many layers of abstraction that have been developed to help manage this complexity. Twotechnology trends highlight the problem.  First, emerging micro-processors (such as Sun's multi-core, multithreaded Niagara) have subtle performance characteristics that are extremely difficult to understand and exploit, even for expert programmers. At the same time, programmers are using languages such as Java, where the code is interpreted by an abstract machine.  The productivity gains that these languages provide are based, in part, on a higher-level abstraction that frees the programmer from the unpleasant details of real hardware. Between the code written by programmers and the instructions executed by the real hardware lie many other layers, including a language-level virtual machine, the operating system, a low-level virtual machine, and a virtual machine monitor. This scenario is quite likely to arise in current web hosting centres.     Our research explores a variety of ways to break down at run-time the barriers imposed by static layers of abstraction. As one specific project, we will explore integrating compiler-based memory management algorithms into a Java JIT compiler, so that the generated code can interact with the JVM's heap management system, which will in turn inform the operating system of candidates for page prefetch and replacement. As a second specific project, we will investigate the use of dynamically-generated execution traces within the operating system code itself (past projects have applied these techniques to user-level code only). Our initial goal will be to support enhanced profiling and performance instrumentation of the OS, followed by dynamically optimizing execution paths within the OS, with the ultimate goal of integrating user-level and system-level execution traces.