Skip to content

Cut the longest CI job from about seven minutes to about two #217

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MarkusPaulsen wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Cut the longest CI job from about seven minutes to about two #217

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 29 additions & 1 deletion pom.xml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,11 @@
<!-- Isolate test classes so WALA state and test-created threads cannot accumulate
in one long-lived JVM. The signed outcome cache preserves cross-fork reuse. -->
<surefire-reuse-forks>false</surefire-reuse-forks>
<!-- The fork heap, as a property for the same reason `surefire-reuse-forks` is one: a
profile that changes how tests are isolated usually has to change how much heap that
isolation needs, and overriding the whole `argLine` to do so would duplicate the
agent and bootclasspath arguments into every profile that tried. -->
<surefire-max-heap>6g</surefire-max-heap>
<jacoco-version>0.8.15</jacoco-version>
<!-- BUNDLE-level minima, shared by the `coverage` profile (one local full-suite
run) and the `coverage-aggregate` profile (the merged CI runs) so that the two
Expand Down Expand Up @@ -501,7 +506,7 @@
the runner dies. 2400s sits above a legitimate full-suite run yet below
the 45-minute CI job timeout, so a real hang is caught with a report. -->
<forkedProcessTimeoutInSeconds>2400</forkedProcessTimeoutInSeconds>
<argLine>@{argLine} -Xmx6g -javaagent:${byte-buddy-path} -Xbootclasspath/a:${aspectj-path} ${jvm.module.access.args}</argLine>
<argLine>@{argLine} -Xmx${surefire-max-heap} -javaagent:${byte-buddy-path} -Xbootclasspath/a:${aspectj-path} ${jvm.module.access.args}</argLine>
</configuration>
</plugin>
<!-- 5. Package -->
Expand Down Expand Up @@ -1038,12 +1043,35 @@
</profile>
<profile>
<id>unit-architecture-tests</id>
<properties>
<!-- One long-lived fork instead of nineteen short ones needs a stated ceiling
rather than the default, because the reason that default exists is the
opposite case. Measured peak resident size of the reused fork at 6g was
3.12 GB and 3.95 GB across two runs, a spread wide enough that the only real
bound was `-Xmx` itself. 3g bounds it deliberately, and three runs at that
ceiling peaked at 1.76, 2.20 and 1.83 GB. -->
<surefire-max-heap>3g</surefire-max-heap>
</properties>
<build>
<plugins>
<plugin>
<groupId>org.apache.maven.plugins</groupId>
<artifactId>maven-surefire-plugin</artifactId>
<configuration>
<!-- These 19 classes share one fork, as `unit-core-tests` and
`integration-core-tests` already do. Without this the profile inherits
`surefire-reuse-forks`, which the `coverage` profile sets to false, and
every class pays a full JVM start: measured at roughly 15.5 s per fork
in CI, against 105 s of actual test time for the whole set.

Whether WALA state survives a shared fork is not settled by argument,
so it was measured: five runs of this profile, all 19 classes and all
188 tests green, peak resident size 1.76 to 2.20 GB at the 3g ceiling.
The integration matrix in maven.yml has run the WALA modes with
`-Dsurefire-reuse-forks=true` all along, which is a related but not
identical case: it selects methods rather than whole classes. If this
profile ever turns flaky, this line is the first thing to revert. -->
<reuseForks>true</reuseForks>
<includes>
<include>**/api/architecture/**/*Test.java</include>
</includes>
Expand Down