7/27/2019 kvm-japan

1/25

KVM: A Hypervisorfor All Seasons

Avi Kivity

avi@qumranet.com

November 2007

7/27/2019 kvm-japan

2/25

Copyright 2007 Qumranet, Inc. All rights reserved.2

Virtualization

Simulation of computer system in software

Components

Processor: register state, instructions, exceptions

Memory Management: paging, protection, tlb

I/O: Networking, storage, human interface

Performance and fidelity are critical

7/27/2019 kvm-japan

3/25

Copyright 2007 Qumranet, Inc. All rights reserved.

Uses

Server consolidation

Testing, R&D

Virtual Desktops

7/27/2019 kvm-japan

4/25

Copyright 2007 Qumranet, Inc. All rights reserved.4

Virtualization basics

Trap changes to privileged stateGuest cannot access hardware

Hide privileged state

Guest cannot detect that the host is changing things behind itsback

Example: interrupt enable flag

7/27/2019 kvm-japan

5/25

Copyright 2007 Qumranet, Inc. All rights reserved.

A Look Back

VMWare

Just-in-time compilation of binaries

Xen

Paravirtualized guests

7/27/2019 kvm-japan

6/25

Copyright 2007 Qumranet, Inc. All rights reserved.6

Virtualization

Simulation of computer system in software

Components

Processor Management

Memory Management

IO Management

Difference from emulation is emphasis on near-nativeperformance

7/27/2019 kvm-japan

7/25

Copyright 2007 Qumranet, Inc. All rights reserved.

The KVM approach

Reuse Linux code as much as possible

Focus on virtualization, leave other things to respectivedevelopers

Integrate well into existing infrastructure, codebase, andmindset

7/27/2019 kvm-japan

8/25

Copyright 2007 Qumranet, Inc. All rights reserved.8

kvm process model

kernel

task task guest task task guest

7/27/2019 kvm-japan

9/25

Copyright 2007 Qumranet, Inc. All rights reserved.9

KVM process model (cont'd)

Guests are scheduled as regular processes

Guests are processes that can execute in three modes:user mode, kernel mode, and guestmode

kill(1), top(1) work as expected

Guest physical memory is mapped into the task's virtualmemory space

Virtual processors in a VM are threads in the hostprocess

7/27/2019 kvm-japan

10/25

Copyright 2007 Qumranet, Inc. All rights reserved.10

KVM Execution Model

Native GuestExecution

Kernelexit handler

Userspaceexit handler

Switch to

Guest Mode

LightweightExit

ioctl()

Userspace Kernel Guest

7/27/2019 kvm-japan

11/25

Copyright 2007 Qumranet, Inc. All rights reserved.11

What's handled in the kernel?

CPU virtualization (special instructions)

MMU virtualization

Local APIC, PIC, and IOAPIC

(planned) paravirtualized network and block device

(planned) paravirtualized guest kernel support code

7/27/2019 kvm-japan

12/25

Copyright 2007 Qumranet, Inc. All rights reserved.12

KVM Initial Design Parameters

x86 only

Focus on full virtualization

Zero modifications to host kernels

7/27/2019 kvm-japan

13/25

Copyright 2007 Qumranet, Inc. All rights reserved.13

We were wrong!

Support for new architectures is underways390: the big daddy of virtualization

PowerPC: for embedded/consumer electronics

ia64: large machines

Paravirtualization supportImproves performance, timing accuracy

Modifications to host to improve scheduling, addswapping

7/27/2019 kvm-japan

14/25

Copyright 2007 Qumranet, Inc. All rights reserved.

Oh yes, swapping

KVM can swap guest memory

Needed for untrusted guests

Enables mmu games like page migration neededfor good NUMA support

7/27/2019 kvm-japan

15/25

Copyright 2007 Qumranet, Inc. All rights reserved.

KVM is a full-spectrum hypervisor

Servers

Desktops/laptops

Small/embedded boards

Really large machines

7/27/2019 kvm-japan

16/25

Copyright 2007 Qumranet, Inc. All rights reserved.

KVM on servers

Same management tools and infrastructure as Linux

Supports virtual machines and regular processesside-by-side

Useful for real-time server processing workloads

Uses the Linux scheduler and I/O stack

Live migration

Integrates with cluster filesystems

7/27/2019 kvm-japan

17/25

Copyright 2007 Qumranet, Inc. All rights reserved.

KVM on desktops/laptops

Will boot on anything Linux boots

Normal desktop doesn't change

Excellent power management

Suspend/resume while virtual machines are running

Swap idle VMs when not in use

Passthrough devices

USB

PCI (planned)

7/27/2019 kvm-japan

18/25

Copyright 2007 Qumranet, Inc. All rights reserved.

KVM for embedded

Board support is in Linux already

Can make footprint as small as needed

Real-time scheduling support

7/27/2019 kvm-japan

19/25

Copyright 2007 Qumranet, Inc. All rights reserved.

KVM and big iron

Leverage many years of Linux scalability workNUMA

Page migration

Linux already supports 4096-processor machines

Competitors can only dream of this

7/27/2019 kvm-japan

20/25

Copyright 2007 Qumranet, Inc. All rights reserved.

Paravirtualized device drivers

virtio

Common drivers for all hypervisors

Hypervisor-specific backend

KVM backend is in progress

7/27/2019 kvm-japan

21/25

Copyright 2007 Qumranet, Inc. All rights reserved.

Release philosophy

Development snapshots every 1-2 weeksRelease early and often

Features introduces quickly

Bugs fixed quickly

Bugs added quickly...Allows developers and users to track and test the latestand greatest

Stable releases part of Linux 2.6.x

With bugfixes going into Linux 2.6.x.y

7/27/2019 kvm-japan

22/25

Copyright 2007 Qumranet, Inc. All rights reserved.

Differences between Xen and KVM

No dom0Removes scaling bottleneck

Removes priority inversion problem

Interrupts handled directly by host

Xen requires the scheduler to dispatch interrupts

Uses the Linux bootstrap

Bringing up KVM on a new machine is trivial

Uses well-known Linux environment

Rich kernel APIsLarge developer base

7/27/2019 kvm-japan

23/25

Copyright 2007 Qumranet, Inc. All rights reserved.

Status & plans

StatusMany supported guests

Good cpu/mmu performance

Good host scaling

PlansAdd more paravirtualization (both cpu and I/O)

Improve guest scaling

Add more hardware support (architectures, hardwarefeatures)

Improve swapping

7/27/2019 kvm-japan

24/25

Copyright 2007 Qumranet, Inc. All rights reserved.

KVM pros

Leverages Linux scheduler, memory management, I/O

No scheduler involvement for I/O

Uses existing Linux security model (can run VM asordinary user)

Uses existing management tools

Power management

Guest memory swapping

Real-time schedulingLeverages Linux development momentum

7/27/2019 kvm-japan

25/25

Thank You