[To be translated] Nova：libvirt image 的生命周期

翻译自：http://www.pixelbeat.org/docs/openstack_libvirt_images/

The main stages of a Virtual Machine disk image as it transfers through OpenStack to be booted under libvirt are:

Initially the image is downloaded from glance and cached in libvirt base. We'll consider the options for handling a qcow2 image stored in glance, as that format can be downloaded quite efficiently from glance as it supports compression, and image sparseness can be maintained. This article will focus on the flow and transformations in "libvirt base", which is used to cache, preprocess and optionally back, VM disk images.

Configuration

First we'll summarize the config variables involved, before presenting the operations associated with each config combination, in each OpenStack release. Note I'm describing upstream OpenStack here, and not my employer's Red Hat OpenStack which has back-ported enhancements between versions where appropriate.

The main reason that raw images are written in "libvirt base" by default (since Diablo), is to remove possible compression from the qcow2 image received from glance. Note compression in qcow2 images is read only, and so this will impact reads from unwritten portions of the qcow2 image. Users may want to change this option, depending on CPU resources and I/O bandwidth available. For example, systems with slower I/O or less space available, may want to trade the higher CPU requirements of compression, to minimize input bandwidth. Note raw images are used unconditionally with libvirt_images_type=lvm.

Whether to use CoW images for the "libvirt instance disks" also depends on I/O characteristics of the user's system. Without CoW, more space will be used for common parts of the disk image, but on the flip side depending on the backing store and host caching, there may be better concurrency achieved by having each VM operate on its own copy.

Enabling preallocation of space for the "libvirt instance disks" can help with both space guarantees and I/O performance. Even when not using CoW instance disks, the copy each VM gets is sparse and so the VM may fail unexpectedly at run time with ENOSPC. By running fallocate(1) on the instance disk images, we immediately and efficiently allocate the space for them in the file system (if supported). Also run time performance should be improved as the file system doesn't have to dynamically allocate blocks at run time, reducing CPU overhead and more importantly file fragmentation.

Disk image operations

For each release and config combination, here are the created files and associated operations in getting a qcow2 image from glance through to being booted in a libvirt Virtual Machine.

Folsom, force_raw_images=True, use_cow_images=True

This results in each instance booting from a CoW image, backed by a resized raw image.

Folsom, force_raw_images=True, use_cow_images=False

This results in each instance booting from a copy of a resized raw image.

Folsom, force_raw_images=False, use_cow_images=False

This results in each instance booting from a copy of a resized qcow2 image.

Folsom, force_raw_images=False, use_cow_images=True

This results in each instance booting from a CoW image, backed by a resized qcow2 image.

Grizzly, force_raw_images=True, use_cow_images=True

Grizzly introduces a change for use_cow_images=True, where it will resize in the $instance_dir rather than in base_. So the resize will not be cached, but this is minimal CPU tradeoff per instance boot, for the extra space saved in base_. We'll just present the default config values here which illustrates the only significant change from Folsom.
This results in each instance booting from a resized CoW image, backed by a raw image.

[² Update Sep 2013: Stanislaw Pitucha also noticed that the above referenced Grizzly change introduced a regression where unpartitioned qcow2 images were no longer resized. See the Havana resize_fs_using_block_device option below for details.]

Grizzly, preallocate_images='space'

Grizzly also has new fallocate functionality in this area controlled by the preallocate_images config option. If that is set to 'space', then after the operations above, the $instance_dir/ images will be fallocated to immediately determine if enough space is available, and to possibly improve VM I/O performance due to ongoing allocation avoidance, and better locality of block allocations.

¹ Havana, resize_fs_using_block_device=False

As noted in the first Grizzly change above, Stanislaw Pitucha noticed that change introduced a regression where unpartitioned qcow2 images were no longer resized. He supplied a fix to resize qcow directly rather than relying on the raw image being available, which would also cater for the force_raw_images=False case that even pre Grizzly did not. This new option can be used to enable this support, but there are some large performance and possible security issues so it's not enabled by default. This support will be available in the upcoming Havana release.

General performance considerations

Performance has improved in this area through each OpenStack release, with some of the main topics to consider, for past and future changes being:

Minimize I/O

Note these were implemented in Essex:

• Copy images, then resize, rather than vice versa
• Directly generating images in the $instance_dir/
• Intelligent reading of sparse input
• Reproduction of sparse input on output
• Use compression

Note this was implemented in Folsom:

• Avoid file sytem overhead by setting libvirt_images_type=lvm. Note file system overhead varies depending on file system

Minimize storage

• Use compression
• Use sparse output/generation
• Avoid resized copies when not needed
• Use CoW if appropriate

Improve caching

• Avoid thrashing the page cache with large intermediate images
• Improve low level caching through better storage allocation

Preprocessing

• Preprocessing may be possible on images like preallocation=metadata which trades off initial CPU cost for possibly much better run time I/O performance
• Such cost would be some what alleviated by having asynchronous population of the base_ cache