Wednesday, March 11, 2015

Virtualization vs. Containerization

Containers provide isolated runtime environments for applications: the entire user space environment is exclusively presented to the container, and any changes to it do not impact other containers' environments. To provide this isolation, a combination of OS-based mechanisms is used: Linux name spaces are used for isolation and scoping mechanism. File system mounts define what files are accessible to the container. cgroups define resource consumption of containers. Still all containers share the same OS kernel which can realize memory footprint efficiencies when identical libraries are used by multiple containers.

With system virtualization, the hypervisor provides a full virtual machine to a guest: the entire OS image including the kernel is now dedicated to the virtual machine. CPU virtualization is used to provide each guest with an exclusive view of a full system environment, and these mechanisms also ensure isolation from other guests. Hypervisor-based management of virtual CPUs, memory and I/O devices is used to define resource consumption of guests.

Which one is better?

As always, it depends on your needs. If you just want to have a number of separate instances to run applications, a container environment often provides greater efficiency, both in managing the application environment, starting the application instances, and in resource consumption. Simple modification and deployment of application environments has been a design principle of container solutions like Docker and is entirely in the DevOps spirit (guess you just have been waiting for more buzzwords).

If you want to have best isolation of environments and come from a server virtualization perspective, then system virtualization may be more relevant: Noisy neighbours are much less of an issue than with containers. While many of the container folks currently focus on improving container isolation, virtual machine isolation is still superior. Coming from physical servers, virtual servers are a natural step, and an existing ecosystem around server management can often be applied to virtual servers, too.

On z systems, Linux has good scalability (to run containers), but z is the platform with an extremely efficient virtualization technology (to run virtual servers), and it is inherent in the entire system architecture. Without having measured it, combining the technologies is probably less painful on z Systems than on other platforms.

There is a third way: both.
Combining system virtualization with containers can be done in multiple ways:

1. A Virtual Machine in a Container

Docker has quite some flexibility on where containers are deployed. One option (called "execution driver") is to use KVM images. This allows to use all the DevOps methods of Docker and combine it with best isolation available, at the cost of having to start up entire Operating System instances when starting containers -- meaning startup time and footprint. Memory efficiencies can only be realized through Kernel Samepage Merging (KSM) -- less effective and efficient, but it's a start.

2. A Container in a Virtual Machine

Conversely, you can run a virtual machine and start a container inside. The VM would not be controlled by Docker, but by existing virtualization management infrastructure. Once the OS instance is up, starting a container would then be done using Docker, and no special setup has to be performed for running containers. Again, containers would have strong isolation since the next container runs in another virtual system; footprint-wise, efficiencies would only be possible through memory deduplication techniques by the hypervisor.

2b. Multiple Containers in a Virtual Machine

This is a variation of running Docker in a VM, suitable for multi-tenancy environments. Here, the assumption is that strongest isolation is only required between containers of different tenants, and straight Linux container isolation is good enough between several containers of the same tenant. Advantages are lower number of virtual machines to run, best isolation between tenants, and enjoying all the efficiencies of vanilla Docker setups.

To summarize: it depends.


  1. Don't understand the intent for last section. Am I missing something. Why would somebody put docker host in a VM. I can imagine we doing this in test on virtual box but in prod I guess does not make sense

    1. Hi Vijay, on z Systems (the background here), virtualization is the default. You cannot even run bare metal, and most Linux customers run under a second level hypervisor. And that includes production workload. This may also be due to the fact that most z shops have a low number of physical mainframe boxes :-). Anyway, on z, this involves very low virtualization overhead.
      However, virtualization takes away some chances for Docker to increase density, therefore a tenant granularity may be the right level to provide good isolation.

  2. "Docker allows development on top of a platform (PaaS) without requiring that platform to be physically present inside the description file.

    However, as has been noted, the Docker technology is not free from problems, a lot of which come from the immaturity of the required container technology in Linux. The specific issues are security (the ability of root to break out of the container environment) and isolation (the ability of one container to steal resources that are needed to service another). While the Dockerized applications run safely behind the enterprise firewall, these issues can largely be ignored. However, one of the promises of Docker is the ability to run anywhere and at some point the Dockerized applications need to leave the safety of the enterprise firewall and provide real services for other businesses or customers."

  3. Vijay, absolutely agree with you. There is no sense to run Docker inside VM on production. The main issues that prevent to run default Docker containers in production mode are security and poor isolation. Plus stateless design (because the lack of live and offline migration) adds even more complexity for enterprise applications during the adoption/migration process to containers architecture.

    However there is a solution to run containers on bare metal hardware. Only in this case the promised density is the best. And less unneeded complexity provides maximum ROI. We are using combination of Docker templates and Virtuozzo containers, and we run containers on top of bare metal hardware since 2012 in production mode. You can find more details at the link

    1. On z Systems, virtualization overhead is not really impacting density, but of course Docker will only be able to provide good density if many containers can be run within one OS instance. Also, even 2nd level virtualization and production is the default on z (for Linux, at least), with very low virtualization overhead. This may be a different story on x86, though.
      You are right on the issues, and they can be addressed through virtualization -- at least on a tenant basis (and very natural so on the mainframe). Agree on the statement about stateless design. Approaches like etcd solve some issues, but for real state, you want stability: backend databases may not need or want the flexibility offered by Docker.


Note: Only a member of this blog may post a comment.