RFR: 8294403: [REDO] make test should report only on executed tests

Aleksey Shipilev shade at openjdk.org
Tue Jan 3 17:30:49 UTC 2023


On Tue, 3 Jan 2023 09:39:59 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> This should help to speed up tests significantly. Currently, if we run "make test" with a subset of tests, JTReg would still read the entirety of test root to report on tests that were not run. Even with current suite of tests it gets expensive. If you add more tests to suite -- for example, mounting a large test archive like pre-generated fuzzer tests -- it starts to take a lot of time.
> 
> The reasonable default for older jtreg is `-report:executed`. My own CI -- that has millions of tests mounted in `test/` -- was running with `JTREG="OPTIONS=-report:executed"` for years now. I think that should be the default way to run the tests in JDK tree.
> 
> [CODETOOLS-7903323](https://bugs.openjdk.org/browse/CODETOOLS-7903323) provides a new reporting option, `-report:files`, which makes this whole business easier. This requires new jtreg.
> 
> Motivational improvements:
> 
> 
> $ time CONF=linux-x86_64-server-release make run-test TEST=gc/epsilon/TestHelloWorld.java
> 
> # Default JDK tree
> 
> # Baseline
> real	0m9.086s
> user	0m22.857s
> sys	0m4.202s
> 
> # Patched
> real	0m6.406s   ; +40% faster
> user	0m17.156s
> sys	0m3.193s
> 
> # +100K fuzzer tests in tree
> 
> # Baseline
> real	3m1.997s
> user	3m16.643s
> sys	0m10.490s
> 
> # Patched
> real	0m8.919s   ; 20x faster
> user	0m17.904s
> sys	0m4.860s
> 
> 
> Additional testing:
>  - [x] Ad-hoc timing tests (see above)
>  - [x] Reproducer from [CODETOOLS-7903331](https://bugs.openjdk.org/browse/CODETOOLS-7903331)
>  - [x] Eyeballing the test results on passing tests with REPEAT_COUNT > 1
>  - [x] Eyeballing the test results on intermittently failing tests with RETRY_COUNT > 1

This might need more exposure.

-------------

PR: https://git.openjdk.org/jdk/pull/11824



More information about the build-dev mailing list