CAREER: Implementing a New Parallelization Model for Virtual Machines

最近的趋势正在迅速改变开发人员和用户对编程语言和应用程序的期望。首先，不断增长的程序复杂性增加了对高级编程语言的需求。第二，互联网的日益增长的可用性和可访问性正在推动对软件的可移植分发格式的需求，同时沿着安全可靠的运行时环境。高级托管语言沿着与它们相关的运行时非常适合于交付未来软件系统所需的特性。托管语言运行时通过在程序执行期间执行许多辅助任务来提供其功能。然而，这些任务通常也会干扰正常的程序执行，造成停顿，并最终导致程序性能下降。由于未来的应用程序需要更昂贵的功能，以提高速度和防止更复杂的软件攻击，执行这种内联辅助任务的开销可能会变得prohibitive.This研究开发了一个新的虚拟机框架，利用多处理器系统提供的机会，有效地并行内联辅助任务，并使主程序线程在运行时不受阻碍的执行。该框架为虚拟机实现了一种新的并行化模型，允许每个辅助任务在自己的线程中执行，并在单独的处理核心上执行。每个辅助线程都是使用程序切片生成的，只计算正确执行所需的状态。因此，辅助线程独立地、并发地执行，并在结果到期之前将它们的结果传送给主线程。该框架消除了由辅助任务引起的停顿，可以加速程序完成，并为未来的程序支持更有益和昂贵的任务。

Recent trends are rapidly shifting developer and user expectations of programming languages and application programs. Firstly, growing program complexity is increasing the demand for higher-level programming languages. Secondly, growing availability and accessibility of the Internet is driving the demand for a portable distribution format for software, along with a safe and secure runtime environment. High-level managed languages along with their associated runtimes are ideally suited to deliver the features required for future software systems. Managed language runtimes provide their features by performing a number of auxiliary tasks during program execution. However, these tasks often also interfere with normal program execution, create stalls, and ultimately result in degrading program performance. As future applications necessitate even more expensive features to improve speed and protect against more sophisticated software attacks, the overhead of performing such inline auxiliary tasks is likely to become prohibitive.This research develops a new virtual machine framework to exploit the opportunities provided by multiprocessor systems to effectively parallelize inline auxiliary tasks, and enable unhindered execution of the main program thread at runtime. The framework implements a novel parallelization model for virtual machines to allow each auxiliary task to execute in its own thread, and on a separate processing core. Each auxiliary thread is generated using program slicing to only calculate the state necessary for its correct execution. Thus, auxiliary threads execute independently, concurrently, and communicate their results to the main thread before they are due. Relieved of the stalls caused by the auxiliary tasks, the framework can allow accelerated program completion, and support even more beneficial and expensive tasks for future programs.