Image Caching

What is Image Caching?

In order to understand what image caching is and why it is beneficial, it helps to be familiar with the process by which an instance is booted from a given base image. When a new instance is created on a compute node, the following general steps are performed by the compute manager in conjunction with the virt driver:

  1. Download the base image from glance

  2. Copy or COW the base image to create a new root disk image for the instance

  3. Boot the instance using the new root disk image

The first step involves downloading the entire base image to the local disk on the compute node, which could involve many gigabytes of network traffic, storage, and many minutes of latency between the start of the boot process and actually running the instance. When the virt driver supports image caching, step #1 above may be skipped if the base image is already present on the compute node. This is most often the case when another instance has been booted on that node from the same base image recently. If present, the download operation can be skipped, which greatly reduces the time-to-boot for the second and subsequent instances that use the same base image, as well as avoids load on the glance server and the network connection.

By default, the compute node will periodically scan the images it has cached, looking for base images that are not used by any instances on the node that are older than a configured lifetime (24 hours by default). Those unused images are deleted from the cache directory until they are needed again.

For more information about configuring image cache behavior, see the documentation for the following configuration options:

Note

Some ephemeral backend drivers may not use or need image caching, or may not behave in the same way as others. For example, when using the rbd backend with the libvirt driver and a shared pool with glance, images are COW’d at the storage level and thus need not be downloaded (and thus cached) at the compute node at all.

Image Caching Resource Accounting

Generally the size of the image cache is not part of the data Nova includes when reporting available or consumed disk space. This means that when nova-compute reports 100G of total disk space, the scheduler will assume that 100G of instances may be placed there. Usually disk is the most plentiful resource and thus the last to be exhausted, so this is often not problematic. However, if many instances are booted from distinct images, all of which need to be cached in addition to the disk space used by the instances themselves, Nova may overcommit the disk unintentionally by failing to consider the size of the image cache.

There are two approaches to addressing this situation:

  1. Mount the image cache as a separate filesystem. This will cause Nova to report the amount of disk space available purely to instances, independent of how much is consumed by the cache. Nova will continue to disregard the size of the image cache and, if the cache space is exhausted, builds will fail. However, available disk space for instances will be correctly reported by nova-compute and accurately considered by the scheduler.

  2. Enable optional reserved disk amount behavior. The configuration workaround workarounds.reserve_disk_resource_for_image_cache will cause nova-compute to periodically update the reserved disk amount to include the statically configured value, as well as the amount currently consumed by the image cache. This will cause the scheduler to see the available disk space decrease as the image cache grows. This is not updated synchronously and thus is not a perfect solution, but should vastly increase the scheduler’s visibility resulting in better decisions. (Note this solution is currently libvirt-specific)

As above, not all backends and virt drivers use image caching, and thus a third option may be to consider alternative infrastructure to eliminate this problem altogether.