Move CI large runners to free runners

[marcoieni]

What

We want to optimize the resources used by our CI.

We already migrated many large runners to free runners by splitting tests into multiple jobs. E.g. rust-lang/rust#134124

How

In the following we analyze what jobs of the auto branch we haven't moved yet and why.

Here's a list of all large runners, together with the jobs that are still using them and their average cost:

• job-aarch64-linux-8c:
  • aarch64-gnu: ci: switch to linux free arm runners rust#135633
  • aarch64-gnu-debug: ci: switch to linux free arm runners rust#135633
  • dist-aarch64-linux: we tried moving it to the free runner but when cache isn't available the runner it's too slow. See zulip.
• job-linux-8c:
  • dist-arm-linux: ci: split dist-arm-linux job rust#141078
    Moving this job to a free runner makes it too slow. We haven't investigated if it's possible to split dist jobs yet. Also, I wonder if there's the same problem as dist-aarch64-linux. EDIT: splitted now.
  • x86_64-fuchsia: this job is now disabled.
    I tried to convert this job to free runners here but it was too slow. This job can't be parallelized because it is just a build.
  • x86_64-gnu-distcheck: ci: convert distcheck to free runner rust#141382
    I tried to convert this job to free runners here but it was too slow. This job can't be parallelized because it only runs one task (distcheck). Maybe it can be parallelized but it's not trivial.
• job-linux-4c-largedisk:
  • dist-powerpc64le-linux: ci: split powerpc64le-linux job rust#141042
    We can move this to free runner if we solve rust CI: explore if it's possible to free more disk space #183
  • x86_64-gnu-debug:
    Same motivation as dist-powerpc64le-linux 👆
• job-linux-16c:
  • dist-x86_64-linux:
    Same motivation as dist-arm-linux.
  • dist-x86_64-linux-alt:
    Same motivation as dist-arm-linux.
• job-windows-8c:
  • x86_64-msvc: CI: split x86_64-msvc job using windows 2025 rust#135632
    There's CI: split x86_64-msvc job rust#133632 to move this to two free runners, however we haven't merged it yet because the x86_64-msvc job is flaky, so we don't want to add more flakiness to CI. To unblock this, we could either fix the flakiness (Tracking Issue for high failure rates on Windows MSVC CI with filesystem errors rust#127883 or implement bootstrap: retry cargo invocations if stderr contains a known pattern rust#134472
  • i686-msvc: CI: split i686-msvc job to two free runners rust#135616
  • i686-mingw: CI: split i686-mingw job to three free runners rust#135813
    There's a proposal to downgrade the target of this job to tier 2, which means we can delete this job entirely. If the proposal isn't accepted, we can try to split the job into two free runners, but we need to address the flakiness of the job first.
  • dist-x86_64-msvc:
    Same motivation as dist-arm-linux

[marcoieni]

Me and JD discussed about this issue.
I would try to:

• remove the job-linux-4c-largedisk large runner
• see if i686-mingw can be downgraded or split (probably into three jobs, because two wasn't enough). Maybe also upgrade to windows 2025 to try remove the flakiness.

For the other jobs, we concluded that for now it's not worth the effort to try to convert them because they probably high effort.

[marcoieni]

We are done with this. We moved many runners and this is an effort we'll keep an eye on. It doesn't need to have an open issue.

For an up to date state of large vs free runners, see the file https://github.com/rust-lang/rust/blob/master/src/ci/github-actions/jobs.yml