Efficient virtualization on embedded power architecture® platforms

Abstract

Power Architecture® processors are popular and widespread on embedded systems, and such platforms are increasingly being used to run virtual machines. While the Power Architecture meets the Popek-and-Goldberg virtualization requirements for traditional trap-and-emulate style virtualization, the performance overhead of virtualization remains high. For example, workloads exhibiting a large amount of kernel activity typically show 3-5x slowdowns over bare-metal. Recent additions to the Linux kernel contain guest and host side paravirtual extensions for Power Architecture platforms. While these extensions improve performance significantly, they are guest-specific, guest-intrusive, and cover only a subset of all possible virtualization optimizations. We present a set of host-side optimizations that achieve comparable performance to the aforementioned paravirtual extensions, on an unmodified guest. Our optimizations are based on adaptive in-place binary translation. Unlike the paravirtual approach, our solution is guest neutral. We implement our ideas in a prototype based on Qemu/KVM. After our modifications, KVM can boot an unmodified Linux guest around 2.5x faster. We contrast our optimization approach with previous similar binary translation based approaches for the x86 architecture; in our experience, each architecture presents a unique set of challenges and optimization opportunities.