search

airlift / airbase

Base POM for Airlift
Apache License 2.0
48 stars 87 forks source link

Only enable check plugins if air.check.skip-* is false #293

Closed nineinchnick closed 2 years ago

nineinchnick commented 2 years ago

Enable check plugins only when matching air.check.skip-* properties are false. This is supposed to avoid any unnecessary initialization overhead. Another minor improvement is that Maven output is not polluted with unused plugins.

I noticed #69 and I've seen similar issues when disabling the dependency plugin so this is the only check plugin that's skipped using configuration.

Tested roughly (only 3 runs of every goal) by running the following goals in the Trino project, on an otherwise idle r5.large instance:

To make tests quicker, I skip the 3 largest modules, this still leaves out 76 modules to be built.

flags='--strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"'
hyperfine --min-runs 3 "./mvnw clean $flags" "./mvnw install $flags" "./mvnw clean install $flags"

Baseline (Trino master - f7226865e8578339b4e56658b5ec116b872a59c6) yields:

Benchmark 1: ./mvnw clean --strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"
  Time (mean ± σ):      4.660 s ±  0.807 s    [User: 8.118 s, System: 0.627 s]
  Range (min … max):    4.112 s …  5.587 s    3 runs

Benchmark 2: ./mvnw clean install --strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"
  Time (mean ± σ):     231.206 s ±  0.780 s    [User: 335.105 s, System: 15.552 s]
  Range (min … max):   230.722 s … 232.106 s    3 runs

Benchmark 3: ./mvnw install --strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"
  Time (mean ± σ):     172.580 s ±  0.755 s    [User: 206.465 s, System: 12.195 s]
  Range (min … max):   171.715 s … 173.102 s    3 runs

After changes - setting airbase to 120-SNAPSHOT in root pom.xml:

Benchmark 1: ./mvnw clean --strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"
  Time (mean ± σ):      4.829 s ±  0.493 s    [User: 8.295 s, System: 0.725 s]
  Range (min … max):    4.430 s …  5.379 s    3 runs

Benchmark 2: ./mvnw clean install --strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"
  Time (mean ± σ):     228.048 s ±  0.667 s    [User: 324.894 s, System: 15.999 s]
  Range (min … max):   227.287 s … 228.532 s    3 runs

Benchmark 3: ./mvnw install --strict-checksums -V -DskipTests -Dmaven.test.skip=true -Dmaven.site.skip=true -Dmaven.source.skip=true -Dmaven.javadoc.skip=true -Dair.check.skip-all -pl "!:trino-server,!:trino-server-rpm,!docs"
  Time (mean ± σ):     169.230 s ±  1.135 s    [User: 202.097 s, System: 12.201 s]
  Range (min … max):   168.043 s … 170.306 s    3 runs

The gain is small (3s, ~1%) but visible.

I'm also attaching flame graphs from the async-profiler, that shows a 2% difference in how much time (samples) is used for the GC and JIT. mvn-profiles.zip

electrum commented 2 years ago

Thanks!