vms-959: native arm64 runners — kill the QEMU-emulation wall on 26 toolchain jobs - #671
Merged
Merged
Conversation
…ll the emulation wall MEASURED root cause (scoping): the ~8-minute toolchain-cluster wall is not "cold rebuilds" — it is cold rebuilds run under ARM64 QEMU EMULATION on x86_64 runners. The identical proof runs ~10x slower emulated: dcl-native 472s vs its native twin dcl-native-x86_64 42s (same script, same assertions). 26 jobs register docker/setup-qemu-action and `docker run --platform linux/arm64`. The repo is PUBLIC, so GitHub's free hosted `ubuntu-24.04-arm` runners are available with zero org/infra setup. On a native arm64 host, `--platform linux/arm64` runs NATIVELY (platform == host, no QEMU in the path) and the setup-qemu step becomes a harmless no-op. So the change is minimal and safe: flip `runs-on: ubuntu-latest` -> `ubuntu-24.04-arm` on exactly the 26 jobs that contain `linux/arm64`. The 58 x86_64/generic jobs are untouched (byte-identical). Coverage is UNCHANGED — these are the aarch64 activation proofs; running them on real arm64 hardware is a strictly better proof than emulating aarch64 on x86_64. Expected: each ~250-470s job -> ~40-90s; the toolchain-cluster wall ~8min -> ~1.5-2min. Follow-up (not here): the now-pointless setup-qemu step in each converted job is a ~10s no-op that can be pruned once this is proven green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This was referenced Aug 20, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The toolchain-cluster CI wall (~8 min) is cold rebuilds run under arm64 QEMU emulation on x86_64 runners. Measured:
dcl-native472s emulated vs its native twindcl-native-x86_6442s — same script, same assertions, ~10× purely from emulation. 26 jobs registerdocker/setup-qemu-action+docker run --platform linux/arm64.Change
The repo is public, so GitHub's free hosted
ubuntu-24.04-armrunners are available with zero org/infra setup. On a native arm64 host,--platform linux/arm64runs natively (no QEMU in the path) andsetup-qemubecomes a harmless no-op. So: flipruns-on: ubuntu-latest→ubuntu-24.04-armon exactly the 26 jobs that containlinux/arm64. The 58 x86_64/generic jobs are untouched (byte-identical).Coverage
Unchanged. These are the aarch64 activation proofs — running them on real arm64 hardware is a strictly better proof than emulating aarch64 on x86_64. This PR's own run measures the speedup (it touches ci.yml → toolchain filter forces the 26 jobs to run).
Expected: each ~250–470s job → ~40–90s; toolchain-cluster wall ~8min → ~1.5–2min.
Notes
setup-qemustep in each converted job is a ~10s no-op, prunable once proven green.build-and-test(~600–790s, ccache'd in vms-fb8: CI tier-split + KVM accel — per-PR loop off the 46m e2e wall #669 — warms over runs) and the VAX cross-compile cluster (~8.8m, separate — tracked under vms-fb8).⚠ Coordinate
Shared infra. Notifying VAX/Alpha/GCC lanes; GCC lane asked to confirm filters before the finer-path-filter branch (that's a separate later PR — this one changes runners only, no filter change). Hold merge for lane ack + green-by-SHA pre-merge run.
🤖 Generated with Claude Code