Basically, Docker uses RunC to provide lightweight and isolated environment for your software. Is it possible to use other runtimes?

Whole Docker nowadays consists of extensible Moby project, one could implement their own environment and run “containers” with other technologies, for example, pure exec or qemu. For last case I mean Kata Containers project - same containers, but virtual machines under the hood.

So I was using two 5$ hosts on DigitalOcean to test how much overhead Kata adds to my services. One to run container, one to generate workload over internal network.

Here’s a simple gist on how to install Docker CE with Kata on Ubuntu server.

Then you would like to run simple service to compare performance, I was using my dummy-service-py, feel free to use it too.

Workload was generated with Siege utility:

siege -c 512 -t 60s http://addr:port/

Results are following:

🐳                RunC        Kata
Availability      100         100       %
Transactions      43274       17548     hits
Response          0.20        1.17      sec
Rate              724.37      293.05    trans/sec
Concurrency       145.08      342.96

  CPU             ~100%       ~100%
  RAM             ~20M        ~150M

It’s safe to say we can see top performance of RunC container here while Kata container (VM) was a little bit oveloaded so performance degraded. If we drop concurrency down, the gap between those technologies wouldn’t be so huge, but still very significant.

Kata containers fail to contest lightning speeds of pure RunC containers, but look like an interesting tool to limit resources between containers in a “hard” way, for example.

And I must say, whole idea is really ironic: containers came to evade VM overhead and provide lightweight isolation, so here’s VMs for your containers engine, just in case. :)