CSR--AES: VT-ASOS: Virtualization Technologies for Application-Specific Operating Systems on Many-Core HPC Systems

众核系统正在成为高性能计算中占主导地位的体系结构范例。 随着内核数量的不断增加，跨应用程序和系统软件堆栈的有效资源分配和可扩展性变得至关重要。 虚拟化可以解决这些挑战，同时提高这些平台的高性能软件开发的生产力。 VT-ASOS项目增强了虚拟化技术，为并行应用提供了可扩展和可定制的硬件执行环境，我们使用了一种模型，其中构成并行应用的任务被封装在它们自己的运行时执行环境中。VT-ASOS由用户级运行时库组成，并结合了一个最小的自定义客户操作系统，该操作系统了解应用程序的要求并仅实现应用程序所需的功能。每个客户内核都运行在增强的虚拟机监视器（VMM）之上，该监视器可以做出资源管理决策，优化硬件资源对任务的分配。VT-ASOS将支持特定于应用程序的系统配置，特定于应用程序的内存管理器和众核感知内存管理策略。 监控基础设施利用硬件和软件计数器来指导基础设施做出的决策。该研究将改进虚拟化技术，减少虚拟化的开销，提高共享密集共享内存多处理器资源的虚拟机之间的性能隔离。 这个项目的成果将是一个有限的和定义良好的可编程模块集，用于将特定于应用程序的策略和机制插入虚拟机。

Many-core systems are emerging as the dominant architectural paradigm inhigh-performance computing. As the number of cores continues to increase,effective resource allocation and scalability across the applicationand system software stacks becomes paramount. Virtualization canaddress these challenges while simultaneously increasing productivity inhigh-performance software development for these platforms. The VT-ASOSproject augments virtualization technologies to provide encapsulatedand customized hardware execution environments for parallel applications.We use a model in which the tasks comprising a parallel applicationare encapsulated in their own runtime execution environment. VT-ASOSconsists of user-level runtime libraries, combined with a minimal, customguest operating system that knows the application''s requirementsand implements only the functionality that the application needs.Each guest kernel runs on top of an enhanced virtual machine monitor(VMM), which makes resource management decisions that optimize theassignment of hardware resources to tasks.VT-ASOS will support application-specific system configuration,application-specific schedulers, and many-core aware memory management policies. A monitoring infrastructure exploits hardware and software counterto guide the decisions made by the infrastructure.The research uses the Xen VMM and Linux as a testbed. This research will improve virtualization technologies, reducingthe overhead of virtualization and improving performance isolationbetween virtual machines sharing the resources of dense shared-memorymultiprocessors. An outcome of this project will be a limited andwell-defined set of programmable modules for plugging application-specificpolicies and mechanisms into virtual machines.