You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@grails.apache.org by jd...@apache.org on 2026/10/14 13:41:13 UTC
(grails-core) branch feat/app-bench-ci updated: Warm the CI app benches past their JIT drift
This is an automated email from the ASF dual-hosted git repository.
jdaugherty pushed a commit to branch feat/app-bench-ci
in repository https://gitbox.apache.org/repos/asf/grails-core.git
The following commit(s) were added to refs/heads/feat/app-bench-ci by this push:
new a3b5aa39df Warm the CI app benches past their JIT drift
a3b5aa39df is described below
commit a3b5aa39dfc295361d0a38761310a16f69831aeb
Author: James Daugherty <jd...@jdresources.net>
AuthorDate: Wed Oct 14 14:41:08 2026 +0100
Warm the CI app benches past their JIT drift
The 2,000 / 2,000 profile still sampled apps that were warming up: the
gsp-layout chunk medians fell from 690 to 490 us across the series and
the interval did not narrow. Measured from a cold start, every example
app keeps getting faster for tens of thousands of requests, and app1
only settles within 5% of its steady latency after about 28,000. Warm up
30,000 requests and sample 10,000, which leaves roughly ±3-7% intervals
and costs about 20 seconds of requests per nested build on a runner.
---
.github/workflows/benchmark.yml | 2 +-
grails-test-examples/benchmarks/README.adoc | 10 +++++++---
2 files changed, 8 insertions(+), 4 deletions(-)
diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
index ff283fbdfd..a8abf1d12e 100644
--- a/.github/workflows/benchmark.yml
+++ b/.github/workflows/benchmark.yml
@@ -295,7 +295,7 @@ jobs:
- name: "Run app-level indy A/B benches"
timeout-minutes: 180
continue-on-error: true
- run: ./gradlew --no-daemon :grails-test-examples-benchmarks:appIndyBench -PappBenchWarmup=2000 -PappBenchSamples=2000 -PappBenchChunks=10 -PappBenchOutDir="$RESULT_DIR" --max-workers=4
+ run: ./gradlew --no-daemon :grails-test-examples-benchmarks:appIndyBench -PappBenchWarmup=30000 -PappBenchSamples=10000 -PappBenchChunks=10 -PappBenchOutDir="$RESULT_DIR" --max-workers=4
- name: "Upload app benchmark artifacts"
if: always()
uses: actions/upload-artifact@v7.0.1
diff --git a/grails-test-examples/benchmarks/README.adoc b/grails-test-examples/benchmarks/README.adoc
index f02ae901db..c1bf1ef5f6 100644
--- a/grails-test-examples/benchmarks/README.adoc
+++ b/grails-test-examples/benchmarks/README.adoc
@@ -86,9 +86,13 @@ Optional properties:
| `appBenchMaxWorkers` | `4` | `--max-workers` for each nested Gradle invocation
|===
-CI on the `performance` label uses warmup 2000 / samples 2000 / chunks 10. Sampling is cheap next to the
-nested Gradle builds around it (2,000 requests of a millisecond or two take a few seconds), while 300 samples
-after 80 warm-up requests left ±4-14% intervals that flipped verdicts between runs. The CPU ruler written next to each
+CI on the `performance` label uses warmup 30000 / samples 10000 / chunks 10. A cold example app keeps getting
+faster for tens of thousands of requests: measured from a cold start, app1 only settled within 5% of its steady
+latency after about 28,000 requests, and a shorter warm-up leaves that drift inside the measured series, where it
+widens the interval and flips verdicts between runs. The requests are cheap next to the nested Gradle builds around
+them, about 20 seconds per build on a CI runner. Even settled, expect intervals of roughly ±3-7%, so a smaller
+difference reads as `no clear change`. The local defaults are a quick smoke profile, not one to draw verdicts from.
+The CPU ruler written next to each
HTTP entry discards at least 50,000 calls of its own before sampling, whatever `appBenchWarmup` says, so the
runner-health check compares two JIT-settled loops rather than two JVMs caught mid-compilation.