ALCF CI bridge holds a GitHub Actions runner for the whole Aurora pipeline
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 35/100
调研方向
Start by reading .github/workflows/alcf.yml and the existing log-fetching step, then compare the push-model examples linked in the issue. The proposal spans GitHub Actions and GitLab CI, including dispatch authentication, status reporting, and cancellation behavior. Done means the bridge exits promptly, completion reports the matching pipeline result, and the final status and logs remain visible; the open design questions need resolution first.
由索引模型根据 Issue 内容生成。
描述
The ALCF bridge (.github/workflows/alcf.yml) keeps an ubuntu-latest runner alive for the entire Aurora pipeline: after triggering the mirror sync it polls GitLab every 30 s for the pipeline to appear (up to 40 min), then every 60 s until it finishes, with a 330 min timeout to cover the PBS queue wait. Since 2026-09-07 that was 81 jobs and ~3,500 runner-minutes, median 27 min per job, and at least one hit the timeout.
Minutes are free for public repos, but:
- JuliaGPU is on the free org plan, so every sleeping bridge job holds one of the 20 concurrent job slots shared by all JuliaGPU repositories;
- Aurora queue time counts against a GitHub Actions timeout, so a busy queue turns into a red check;
- adopting "the newest pipeline for this SHA" is fragile, which is what needed the re-run workaround in #648.
How others do it
CUDA.jl's benchmark workflow used to wait on Buildkite the same way, and switched to a push model (JuliaGPU/CUDA.jl@7934bfd7): the last Buildkite step calls POST /repos/JuliaGPU/CUDA.jl/dispatches with a benchmarks_complete event, and Benchmark.yml runs on: repository_dispatch, only once results exist. Oceananigans does the same (pipeline-benchmarks.yml). AMDGPU.jl and Metal.jl still poll Buildkite from their Benchmark.yml (via EnricoMi/download-buildkite-artifact-action), and Metal already ran into Buildkite's API rate limit doing so.
Proposal
- Keep the trigger side in GitHub Actions (mirror sync, label-gated fork push, fresh pipeline via trigger token on re-run), but have it exit right away after posting a
pendingcommit status on the head SHA, linking to the GitLab pipeline. - Add a final GitLab job (
when: always, on the login-node shell runner) that sends arepository_dispatch(e.g.alcf_complete, with pipeline id, SHA and status), using a fine-grained PAT scoped to oneAPI.jl stored as a masked GitLab CI/CD variable. - A short workflow
on: repository_dispatchfetches the GitLab job logs (as the current log step does, since the instance requires an ALCF login) and sets the final status/check run on that SHA, linking to its own run.
This frees the runner, decouples Aurora queue time from GitHub timeouts, and matches results by pipeline id instead of guessing. Since the token lives on the GitLab side, fork PRs get results too (unlike CUDA.jl, where fork PRs can't decrypt the dispatch token).
Things to sort out:
- a pipeline that is cancelled (or whose runner dies) never runs the final job, so its status would stay
pending; either accept that or add a scheduled sweep; - re-running the GitHub check would now mean triggering a new pipeline, close to what #648 already does;
- whether ALCF allows outbound HTTPS to
api.github.comfrom CI jobs (the codecov upload suggests so) and storing a GitHub PAT in GitLab variables; - if the ALCF GitLab is Premium, its built-in GitHub integration can post pipeline statuses natively, though that wouldn't solve log visibility.
cc @michel2323
- 主要语言
- Julia
- 星标
- 217
- 派生
- 37
- 平均合并
- 17 小时 9 分钟
- 30 天内合并 PR
- 28
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
JuliaGPU/oneAPI.jl 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
难度 4/5 3-5 天 新手友好度 30/100
维护者通常 1 天内回复
-
难度 4/5 3-5 天 新手友好度 48/100
维护者通常 1 天内回复
-
难度 5/5 一周以上 新手友好度 45/100
维护者通常 1 天内回复
-
难度 4/5 3-5 天 新手友好度 35/100
JuliaGPU/oneAPI.jl#607 · 1 条评论 ·
维护者通常 1 天内回复
查看 JuliaGPU/oneAPI.jl 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 70/100
SciML/DiffEqNoiseProcess.jl#342 ·
-
难度 1/5 1 小时以内 新手友好度 88/100
-
难度 2/5 1-3 小时 新手友好度 72/100
oxfordcontrol/COSMO.jl#211 ·
-
documentation
难度 2/5 半天 新手友好度 65/100
维护者通常 6 天内回复
-
Out-of-place JLArray/GPU problem with VectorContinuousCallback scalar-indexes (callback cache built with CPU zeros)可能已有人在做 @ChrisRackauckas-Claude 于 1 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 74/100
SciML/OrdinaryDiffEq.jl#4813 ·
维护者通常 1 天内回复