Why your QEMU arm64 build is 40× slower

Adding linux/arm64 to a Docker build is one line. The consequence is sometimes a 3× slowdown and sometimes a 40× one, and the difference confuses people. It is not random.

The mechanism

QEMU in user-mode emulation translates arm64 instructions into x64 instructions at runtime. Translation has fixed overhead per instruction. So the penalty scales with how many instructions your build executes — not with how long it takes, and not with image size.

A build that copies files and runs apt-get executes relatively few instructions. A build that runs a compiler executes an enormous number.

Measured

The pattern is clean: the more your build compiles, the worse emulation gets. If your Dockerfile invokes a compiler, QEMU is not a viable strategy.

The fix

Build each architecture on its own hardware and merge the digests into a manifest list. It is more YAML than a single platforms line, and it is the difference between three minutes and two hours.

strategy:
  matrix:
    include:
      - platform: linux/amd64
        runner: runnerhut-8vcpu-ubuntu-2404
      - platform: linux/arm64
        runner: runnerhut-8vcpu-ubuntu-2404-arm
runs-on: ${{ matrix.runner }}