vms-fb8: CI tier-split + KVM accel — per-PR loop off the 46m e2e wall - #669
Merged
Merged
Conversation
… e2e wall Measured (run 32317162722, 46.0m): the wall is one job; the critical path is ~11 heavy QEMU release/install/upgrade/boot-scenario e2e gates (14-46m each), each = cut-release ~19m + image build + boot/verify ~25m under TCG software emulation. 79 of 90 jobs already finish under 10m. C (tier-split): the 12 heavy e2e gates now run on NON-pull_request events only (push, merge_group, workflow_dispatch, schedule) via `if: github.event_name != 'pull_request'`. A PR runs the fast tier (build+test, link/activate, kernel-executive shards, static analysis, conformance, cross-compiles). No coverage deleted — the heavy tier MOVES off the per-PR critical path. A (KVM): run-qemu.sh / run_tests.sh / release_install_inner.sh select `-accel kvm -cpu host` when /dev/kvm is writable, else `-accel tcg` (identical behavior, no regression without KVM). GitHub-hosted Linux runners expose /dev/kvm; accelerates the kernel-executive shards on the PR path and every e2e job on the merge/dispatch path (~10x, per run_tests.sh's own TCG note). Coverage-before-merge preserved: main has no merge queue today, so the reap gate MUST fire a `workflow_dispatch` full run on the candidate branch and verify GREEN before squash-merge (KVM-accelerated ~single-digit minutes). merge_group catches it automatically once a queue is enabled (operator call). Rationale + honest target split + follow-ons (D1 changed-file lint, D2 ccache, B build-once artifact, B' self-hosted KVM) in docs/design-ci-latency-40m-to-1m.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… review) Per gcc-oracle-lane review of #669: the tier-split widens the set of gates that run only on non-PR events (heavy e2e + the pre-existing facility_defects negctl meta-checks), so a lane reaping on fast-tier-green alone gets bitten. Name the pre-merge workflow_dispatch full run as the backstop in the header. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Contributor
Author
|
Reviewed by main/ACP conductor — sound, own CI green. Coverage mechanism holds: heavy jobs now run unconditionally on non-PR events (stronger than the old path-filter skip); workflow_dispatch fires all of them; coverage-before-merge = the reap gate's green-by-SHA Ordering constraint: land AFTER the 0.5 atomic-flip (#655) reaps — the flip's final full-suite proof should run under the CI semantics it was verified against; don't swap per-PR event handling mid-reap. Sequence: reap flip → land #669 → all lanes adopt the pre-merge workflow_dispatch discipline. Ready to merge the instant the flip lands. |
… builds Measured on the PR run: the fast-tier wall is set by the single slowest job on a fully-parallel pool (peak concurrency 58/61 -- no queueing). The slow jobs are cold rebuilds of the same tree (DCL link+activate = 467s in one build step) and static-analysis's whole-tree clang-tidy (10.3m, informational-only). - ccache on build-and-test + static-analysis: keys on preprocessed content, so across runs unchanged TUs are served from cache and only changed files recompile. Populates this run, hits the next -- lowers the build floor (~7m cold -> ~2-3m warm), which is the wall on a parallel pool. - static-analysis clang-tidy: on a PR, lint only files changed vs the base branch (whole-tree still on push/merge_group/schedule). The sweep re-flagged hundreds of untouched files every PR; it is informational-only so this narrows latency, not a gate. Next structural lever (child vms-809, expanded): a single build-native artifact that the ~30 link/activate/probe jobs consume instead of each cold-rebuilding the toolchain -- plus finer path filters so a normal PR runs the jobs its changed files affect, not all 61. Floor is one warm build of the tree. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This was referenced Aug 20, 2026
baron-3dl
added a commit
that referenced
this pull request
Aug 20, 2026
…f the PR loop) (#674) Investigation: netbsd-vax-simh is 312s of un-accelerable SIMH TCG boot (VAX has no KVM) of STOCK NetBSD/vax -- the smoke never loads OVMX source, yet it gated on the broad netbsd_vax filter (covers src/**), so any OVMX-source PR paid 312s for a boot that proves nothing about that source. Move it to NON-pull_request events (push/merge_group/schedule) -- same treatment as the heavy QEMU e2e in #669. The ILP32 width proof stays per-PR via the elf32-vax cross-compile jobs (netbsd-vax-vms/vmsfs/facility-tools-crosscompile, vmsfs-core-elf32-vax); the boot still gates pre-merge + push + nightly. No coverage deleted. 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
Takes the per-PR iteration loop off the 46-minute e2e wall for the 5-wide peer swarm.
Measured (run 32317162722, 46.0m core PR): the wall is one job; the critical path is ~11 heavy QEMU release/install/upgrade/boot-scenario e2e gates (14–46m each), each =
cut-release~19m + image build + boot/verify ~25m under TCG software emulation. 79 of 90 jobs already finish under 10m.Changes
pull_requestevents only (push,merge_group,workflow_dispatch,schedule). A PR runs the fast tier (build+test, link/activate, kernel-executive shards, static analysis, conformance, cross-compiles). No coverage deleted — the heavy tier moves off the per-PR critical path. This PR's own run is a live demo: the heavy jobs should show skipped.run-qemu.sh/run_tests.sh/release_install_inner.shselect-accel kvm -cpu hostwhen/dev/kvmis writable, else-accel tcg(identical behavior, no regression without KVM). Accelerates the kernel-executive shards on the PR path + every e2e job on the merge/dispatch path (~10×, perrun_tests.sh's own TCG note).Coverage-before-merge (important)
mainhas no merge queue today, somerge_groupnever fires. Until a queue is enabled, the reap gate must fire aworkflow_dispatchfull run on the candidate branch and verify GREEN before squash-merge (KVM-accelerated, single-digit minutes). Documented in theci.ymlheader. ⚠ Operator call: enabling a GitHub merge queue onmainmakes this mechanical.Honest target
Per-PR: heavy e2e wall gone now (~2–4m ceiling: static-analysis + build-and-test); ~1–2m after the follow-ons. Full merge gate: ~4–6m. Literal <60s per-PR is bounded by ctest/clang-tidy — chased in follow-ons (D1 changed-file lint, D2 ccache, B build-once artifact), tracked under vms-fb8.
Rationale + full analysis:
docs/design-ci-latency-40m-to-1m.md.⚠ Coordinate before merge
Shared infra — re-triggers all four lanes. Notifying VAX / Alpha / GCC-oracle peers; the VAX co-release gate (
cut-release-vax-gate) moves to the non-PR tier, so VAX build breakage is caught at the pre-merge full run rather than per-PR. Hold merge for peer ack.🤖 Generated with Claude Code