SHF: Small: Virtualization of Heterogeneous and Non-Uniform Memory Hierarchy

硬件存储设备是分层结构的，并且日益异构性(在易失性、能效和可靠性方面)和非一致性(在延迟和带宽方面)。然而，现有的内存虚拟化技术只为每个虚拟机提供了一个扁平、同构、统一的内存空间，这是对硬件内存资源的不透明和扭曲的抽象。这种抽象给虚拟机中的应用程序提出了有效利用内存资源的巨大挑战。云计算平台和其他虚拟平台构成了工业、军事、学术界和更广泛人群使用的现代计算基础设施的主干。该项目大幅改进了内存虚拟化技术，使这些平台上的应用程序能够充分利用硬件内存资源的异构性、非均一性等特性，提高性能、能效和可靠性。该项目让本科生和研究生参与虚拟化和云的工作。这将为他们提供强大的编程和系统能力，以便在云计算行业工作。该项目探索新的内存虚拟化技术，为每个虚拟机提供由虚拟缓存设备和异类虚拟内存设备组成的完整虚拟内存层次结构。虚拟存储器层次结构有效地用作虚拟机和虚拟机监视器之间的表现性接口。对于一个虚拟机来说，它真实地反映了底层硬件内存资源的功能和体系结构，在此基础上可以有效地应用内存感知和缓存感知优化技术。对于虚拟机监控器来说，虚拟设备的使用反映了虚拟机上的工作负载对不同类型内存资源的需求，在此基础上可以很好地协调内存资源的共享。研究的重点是：1)能够在不同类型的硬件存储设备之上构建虚拟设备并以低成本维护其特征和结构的软件框架和关键硬件支持；2)在虚拟机之间高效共享不同类型的存储资源的技术；3)验证现有的虚拟存储层次结构上的缓存感知和内存感知技术。为了促进成果的传播，预期的框架和技术将被内置到Linux操作系统和基于内核的虚拟机(KVM)等虚拟机监视器中，并向社区开放共享。

Hardware memory devices are hierarchically structured and increasingly heterogeneous (in terms of volatility, energy efficiency, and reliability) and non-uniform (in terms of latency and bandwidth). However, existing memory virtualization technology only provides each virtual machine with a flat, homogeneous, and uniform memory space, which is an opaque and distorted abstraction of hardware memory resources. This abstraction poses great challenges for applications in virtual machines to make efficient utilization of memory resources. Cloud computing platforms and other virtualized platforms form the backbone of the modern computing infrastructures used by industry, military, academia, and the wider population. The project substantially improves memory virtualization technology, so that applications on these platforms can fully leverage the heterogeneity, non-uniformity, and other features of hardware memory resources to improve performance, energy efficiency, and reliability. The project involves undergraduate and graduate students to work with virtualization and the cloud. This will provide them with strong programming and system capabilities to work in the cloud computing industry.The project explores novel memory virtualization techniques to provide each virtual machine with a complete virtual memory hierarchy that consists of virtual cache devices and heterogeneous virtual memory devices. The virtual memory hierarchy effectively serves as an expressive interface between a virtual machine and the virtual machine monitor. For a virtual machine, it faithfully reflects the features and architectures of underlying hardware memory resources, based on which memory-aware and cache-aware optimization techniques can be effectively applied. For the virtual machine monitor, the usage of virtual devices reflects the demand of the workload on a virtual machine for different types of memory resources, based on which the sharing of memory resources can be well coordinated. The investigation focuses on 1) the software framework and critical hardware support that can build virtual devices in virtual memory hierarchy on top of diverse types of hardware memory devices and maintain their features and structures with low cost; 2) the techniques to efficiently share different types of memory resources between virtual machines; and 3) validating existing cache-aware and memory-aware techniques on virtual memory hierarchy. To spur the dissemination of results, the expected framework and techniques will be built into Linux OS and virtual machine monitor such as Kernel-based Virtual Machine (KVM), and be open to the community for sharing.