posted 4 years ago by Ben Cordero

Modern cloud computing doesn't install the same way that "bare-metal" and traditional virtualisation system use. As I have discussed before, they may not even be using a boot loader. This has a dramatic effect on the way cloud servers (aka. instances) are booted.

A first principles approach to converting from traditional systems to the cloud, is to take a raw disk image, and use that in the virtual environment. OpenStack lets you do this, and it will work but there are a few things you can do to improve performance.


Use paravirtualization drivers. Commonly known as virtio-[something], these will offer better performance than letting the hypervisor emulate standard hardware.

Use cloud-initialization routines. The ubuntu cloud-init package (ported to Fedora, Debian and SUSE derivatives, emulated by other packages by other distros) is a good way to individualise a cloud instance once it has booted to make adjustments from the shared cloud image. Such changes include using deploy-time ssh keys (and possibly other security tokens) instead of relying on a shared password (or other secrets), configuration management hooks (e.g. puppet or chef) etc.

Another feature of cloud elasticity is the ability for services to scale vertically. Just ask for more resources from the virtual environment. Giving more CPU cores or more RAM doesn't require sending an engineer to open cases and adding chips. Similarly, expanding HDD space isn't a hands-on task, but requires some tweaks.

Image format

I have already mentioned raw disk images. From traditional metal servers these are easily imported to into OpenStack, but suffer performance issues on common cloud operations, booting, snapshotting, resizing etc. require that the images are copied in their entirety as the hypervisor (and related tooling) remains agnostic to the bytes.

Using qcow2 images (from the qemu suite), a Copy-on-Write format is a virtualisation efficient way to store images. It supports compression and encryption and has a happy ability to use read-only backing stores and uses a separate file for changes. If many instances use the same image, then they can all use the same read-only starting point.

Elasticity, the ability to grow and shrink filesystems in the cloud, is provided by another technique that is realised by the use of cloud systems. Doing away with the bootloader, means that we can avoid using MBR and related structures. The most prominent being fixed sized Partitions.

The AMI format

The Amazon AWS cloud is primarily based on the Xen PV system. Kernel and ramdisk are already outside of the filesystem as AKIs and ARIs (I'll explain those in a bit). The meat of an Amazon image is the Amazon Machine Image, the AMI.

From a bits and bytes point of view, an AMI is the literal filesystem. I don't think this is actually documented anywhere, but that is all there is to it. An AMI is the raw representation of a root filesystem (typically ext4).

A cloud environment such as AWS or OpenStack can use an AMI, combined with an AKI and optional ARI to efficiently create a cloud instance. This boot method, with nomenclature to remind us who named it, I will call The AMI boot method.

The AMI boot method

Step 1: Grab the AMI (a root filesystem) and apply it to a disk/block device to be given to the hypervisor.

Step 2: Resize it to the flavor (typically sans u, blame American centric developers), say 20G. This is the important step unique to the AMI format, since there's no partition information in the user provided image, resizing the filesystem is as easy as resizing the filesystem.

Step 3: Use the metadata stored with the AMI to apply the AKI and ARI (stored separately).

Step 4: Let the hypervisor (Xen, KVM/qemu etc) can now go wild.
Other hypervisors might inject the kernel/initrd into the filesystem, add some bios boiler plate and boot it emulating the traditional process. I haven't checked, but that is certainly possible.

OpenStack takes a further optimization (other clouds might do this too). While the AMI is transported and handled by the user as a raw representation of a filesystem (i.e. you can loop mount the bytes), glance stores and manipulates AMIs using qcow2, so you get all of the goodies such as quick copies, compression etc transparantly.

Creating AMIs

Create a block device, and put a filesystem on it.

# lvcreate vg -n my_ami -L 10G  
# mkfs.ext4 /dev/vg/my_ami  
# mkdir /mnt/amiroot && mount /dev/vg/my_ami /mnt/amiroot

Alternatively, use a loopback device

# qemu-img create -f raw my_ami.img 10G  
# mkfs.ext4 my_ami.img # No partitioning, no offsets involved  
# mkdir /mnt/amiroot && mount -o loop my_ami.img /mnt/amiroot

Curate your linux rootfs under /mnt/amiroot using methods that I have already discussed in this blog. Perhaps by stage3 install, debootstrap or even rsync from a live server. Now is a good time to install cloud-init, enable the ttyS0 console and do other tasks that you want all instances based on this image to have.

At this point, you might want to also add a kernel to the filesystem using the distro's package manager. But you can save some space if you have a cloud- ready kernel/ramdisk prepared.

The above commands created a filesystem that was 10G in size. Stored raw, this is a bit unwieldy (even with filesystem holes) since uploads of this image will send the literal zeroes.

# AMI_IMG=/dev/vg/my_ami or AMI_IMG=my_ami.img  
# e2fsck -f $AMI_IMAGE  
# resize2fs -M $AMI_IMAGE  
# BLOCK_COUNT=$(tune2fs -l $AMI_IMG|awk '/Block count:/ {print $3}')

Fsck needs to run prior to any resizing efforts. The resize itself uses the -M flag, which will shrink the filesystem automatically without the user needing to guess how small it needs to be. The filesystem's size can be retrieved using tune2fs, we store it in $BLOCK_COUNT.

By default, ext4 will use a block size of 4096 bytes per block. Thus the filesystem size is 4096 * $BLOCK_COUNT.

Uploading AMIs

Start with the AKI and ARI.

# KERNEL_ID=$(glance image-create \  
 --name="my_aki" \  
 --disk-format=aki \  
 --container-format=aki \  
 < boot/vmlinuz-* \  
 | awk '/ id / { print $4 }')  
# INITRD_ID=$(glance image-create \  
 --name="my_ari" \  
 --disk-format=ari \  
 --container-format=ari \  
 < boot/initrd-* \  
 | awk '/ id / { print $4 }')

Adapt the command to your needs. The AKI and ARI are real kernels and initramfses, the real ouput from a kernel build/install and might already be compressed.

To link the kernel and ramdisk to the main image, we save their UUIDs and add them to the AMI metadata.

# dd if=$AMI_IMG bs=4096 count=$BLOCK_COUNT \  
>    | glance image-create \  
>    --name="my_ami" \  
>    --disk-format=ami \  
>    --container-format=ami \  
>    --property kernel_id=${KERNEL_ID} \  
>    --property ramdisk_id=${INITRD_ID}

We only upload the necessary bytes to glance. Typically only a few hundred MB, not the full 10G.

Recent Posts


Atom / RSS