CSR: Medium: Multicore Virtual Machines for High-Speed Networking

在现代计算机系统中，新兴的多核芯片多处理器（cmp）提供的计算资源与虚拟机（VM）（如Java VM或Microsoft公共语言运行时）利用这些资源的能力之间存在着不匹配。CMP架构将主导未来几代处理器；同时，出于安全性和可移植性的考虑，软件将针对虚拟机而不是本地处理器。然而，由于现代vm无法直接利用cmp提供的粗粒度并行性，因此它们在未来几代处理器中不会获得显著的性能提升。这项工作旨在开创一种方法，通过创建流水线虚拟机来加速cmp上vm的执行，流水线虚拟机实现为一系列平衡的程序，或流水线阶段，每个在不同的处理器核心上执行。具有N个平衡级的流水线虚拟机相对于非流水线虚拟机可以提供N倍的性能提升。流水线虚拟机允许程序员在编写顺序单线程程序时受益于多核并行性。本研究议程紧密结合了虚拟机流水线的优点和局限性的理论工作，包括1)实际系统上流水线的设计和实现，以及2)探索新的计算机体系结构和机制。具体来说，研究计划的重点是针对高速网络的流水线虚拟机，在这个领域，激进的cmp已经部署在商业和学术平台上，并且公共可访问的测试平台将允许其他人开发和部署他们自己设计的流水线虚拟机。

A mismatch exists in modern computer systems between the computational resources provided by emerging multicore chip-multiprocessors (CMPs) and the ability of virtual machines (VMs), such as the Java VM or Microsoft Common Language Runtime, to exploit those resources. CMP architectures will dominate the next several processor generations; meanwhile, for reasons of security and portability, software will target VMs rather than native processors. However, since modern VMs are ill-equipped to directly exploit the coarse-grained parallelism provided by CMPs, they will not experience significant performance increases with future generations of processors.This work aims to pioneer an approach for accelerating the execution of VMs on CMPs by creating pipelined VMs which are implemented as a series of balanced programs, or pipeline stages, each executing on a distinct processor core. A pipelined VM with N balanced stages can provide an N-fold increase in performance relative to an unpipelined VM. Pipelined VMs allow programmers to benefit from multicore parallelism while writing sequential single-threaded programs.This research agenda closely integrates theoretical work on the benefits and limits of VM pipelining, with both 1) the design and implementation of pipelines on real systems, and 2) the exploration of novel computer architectures and mechanisms. For concreteness, the research plan focuses on pipelined VMs that target high-speed networking, an area in which aggressive CMPs are already deployed in commercial and academic platforms and where publically-accessible testbeds will allow others to develop and deploy pipelined VMs of their own design.