Theses Doctoral

The Design, Implementation, and Evaluation of Software and Architectural Support for Nested Virtualization on Modern Architectures

Lim, Jin Tack

Nested virtualization, the discipline of running virtual machines inside other virtual machines, is increasingly important because of the need to deploy workloads that are already using virtualization on top of virtualized cloud infrastructures. However, nested virtualization performance on modern computer architectures is far from native execution speed, which remains a key impediment to further adoption. My thesis is that simple changes to hardware, software, and virtual machine configuration that are transparent to nested virtual machines can provide near-native execution speed for real application workloads. This dissertation presents three mechanisms that improve nested virtualization performance.

First, we present NEsted Virtualization Extensions for Arm (NEVE). As Arm servers make inroads in cloud infrastructure deployments, supporting nested virtualization on Arm is a key requirement. The requirement has recently been met with the introduction of nested virtualization support for the Arm architecture. We built the first hypervisor using Arm nested virtualization support and show that, despite similarities between Arm and x86 nested virtualization support, performance on Arm is much worse than on x86. This is due to excessive traps to the hypervisor caused by differences in non-nested virtualization support. To address this problem, we introduce a novel paravirtualization technique to rapidly prototype architectural changes for virtualization and evaluate their performance impact using existing hardware. Using this technique, we introduce NEVE, a set of simple architectural changes to Arm that can be used by software to coalesce and defer traps by logging the results of hypervisor instructions until the results are actually needed by the hypervisor. We show that NEVE allows hypervisors running real application workloads to provide an order of magnitude improvement in performance over current Arm nested virtualization support and up to three times less overhead than x86 nested virtualization. NEVE is included in the Armv8.4 architecture.

Second, we introduce virtual-passthrough, a new approach for providing virtual I/O devices for nested virtualization without the intervention of multiple levels of hypervisors. Virtual-passthrough preserves I/O interposition while addressing the performance problem of I/O intensive workloads as they perform many times worse with nested virtualization than without virtualization. With virtual-passthrough, virtual devices provided by a host hypervisor, the hypervisor that runs directly on the hardware, can be assigned to nested virtual machines directly without delivering data and control through multiple layers of hypervisors. The approach leverages the existing direct device assignment mechanism and implementation, so it only requires virtual machine configuration changes. Virtual-passthrough is platform-agnostic and easily supports important virtualization features such as migration. We have applied virtual-passthrough in the Linux KVM hypervisor for both x86 and Arm hardware, and show that it can provide more than an order of magnitude improvement in performance over current KVM virtual device support on real application workloads.

Third, we introduce Direct Virtual Hardware (DVH), a new approach that enables a host hypervisor to directly provide virtual hardware to nested virtual machines without the intervention of multiple levels of hypervisors. DVH is a generalization of virtual-passthrough and does not limit virtual hardware to I/O devices. Beyond virtual-passthrough, we introduce three additional DVH mechanisms: virtual timers, virtual inter-processor interrupts, and virtual idle. DVH provides virtual hardware for these mechanisms that mimics the underlying hardware and, in some cases, adds new enhancements that leverage the flexibility of software without the need for matching physical hardware support. We have implemented DVH in KVM. Our experimental results show that combining the four DVH mechanisms can provide even greater performance than virtual-passthrough alone and provide near-native execution speeds on real application workloads.

Files

  • thumnail for Lim_columbia_0054D_16278.pdf Lim_columbia_0054D_16278.pdf application/pdf 842 KB Download File

More About This Work

Academic Units
Computer Science
Thesis Advisors
Nieh, Jason
Degree
D.E.S., Columbia University
Published Here
November 16, 2020