Benchmark containerd Runtimes

This repo contains a Go program that allows you to perform benchmarks on different low-level container runtimes that are OCI-compliant. It uses the containerd client to spawn a configurable amount of containers.

Benchmarks

Current supported benchmark types are startup times and memory usage.

Results

There are benchmark run results available in the results directory. These results are startup and memory benchmarks for runc, Kata Containers (Kata 2.x & Kata 3.x, using different hypervisors (QEMU, Firecracker, Cloud Hypervisor, Dragonball)), and gVisor (KVM & systrap platforms).

Building

You can build the binaries in the benchmarks module as follows:

Startup

CGO_ENABLED=0 go build ./cmd/startup/

Memory

CGO_ENABLED=0 go build ./cmd/memory/

Usage

# Replace [benchmark] with either `startup` or `memory`
❯ ./[benchmark] --help
Usage of ./[benchmark]:
  -iterations int
        Amount of benchmark iterations (default 10)
  -meta string
        Additional metadata. Can for example be '-firecracker' to indicate firecracker VMM.
  -network-gateway string
        The gateway of the CNI network. This can be different depending on CNI configuration (default "10.4.0.1")
  -runtime string
        The runtime to run the benchmark for (default "io.containerd.runc.v2")
  -snapshotter string
        The snapshotter that should be used. For Firecracker this should probably be 'devmapper' (default "overlayfs")