| commit | 9adbe0f3b1c725419e7cc8c40a6c393a8ae86009 | [log] [tgz] |
|---|---|---|
| author | Googler <leba@google.com> | Tue Mar 14 09:17:35 2023 -0700 |
| committer | Copybara-Service <copybara-worker@google.com> | Tue Mar 14 09:18:39 2023 -0700 |
| tree | 9752b0e009cf4814faa3370f8c35ff43c8b66db7 | |
| parent | 08d31877bd5c1500e3ca67f47bcd3b79b4af226d [diff] |
[Skymeld] Introducing `--experimental_skymeld_analysis_overlap_percentage`. This flag allows us to influence the degree of overlapping between analysis-execution in blaze. 1. Performance Our benchmarks showed that allowing analysis and execution to overlap freely does not always mean improve performance: - Increased resource contention: more inflight evaluations means more states to keep and more contention. - Prolonged analysis phase: the analysis work in Skymeld mode generally finishes much later into the build. This means that we have to hold on to analysis-specific states for much longer. This negatively affects build performance (smaller available heap space -> more GC, increased risk of OOM). The original goal of Skymeld is to make use of the available CPU resources observed _towards the end_ of the analysis phase. With `--experimental_skymeld_analysis_overlap_percentage=5`, we would allow blaze to finish (roughly) 95% of the analysis phase work [1] before letting the execution work begin. 2. A simulation of noskymeld mode. With the flag value set to 0, we can enforce that there's _no_ overlapping between analysis/execution. This would allow us to simulate the noskymeld behavior with skymeld enabled. This can be used to e.g. investigate inefficiencies with Skymeld implementation, and may be a way to get around blockers (e.g. multiple package roots). [1] "roughly" because it's 95% of the top level target, but individually they may take different amount of time to finish. PiperOrigin-RevId: 516543887 Change-Id: I51cdb3dca9221c0d7187942ab8705aa27d1f4963
{Fast, Correct} - Choose two
Build and test software of any size, quickly and reliably.
Speed up your builds and tests: Bazel rebuilds only what is necessary. With advanced local and distributed caching, optimized dependency analysis and parallel execution, you get fast and incremental builds.
One tool, multiple languages: Build and test Java, C++, Android, iOS, Go, and a wide variety of other language platforms. Bazel runs on Windows, macOS, and Linux.
Scalable: Bazel helps you scale your organization, codebase, and continuous integration solution. It handles codebases of any size, in multiple repositories or a huge monorepo.
Extensible to your needs: Easily add support for new languages and platforms with Bazel's familiar extension language. Share and re-use language rules written by the growing Bazel community.
Follow our tutorials:
To report a security issue, please email security@bazel.build with a description of the issue, the steps you took to create the issue, affected versions, and, if known, mitigations for the issue. Our vulnerability management team will respond within 3 working days of your email. If the issue is confirmed as a vulnerability, we will open a Security Advisory. This project follows a 90 day disclosure timeline.
See CONTRIBUTING.md
