bench(gradle): add JVM flags for heap/GC stability

Fixed heap size (-Xms2g -Xmx2g), pre-touch pages, serial GC, and
disabled adaptive size policy to reduce measurement noise from heap
resizes, page faults, and concurrent GC threads.

Author: not-matthias

examples/example-gradle/build.gradle.kts

@@ -11,17 +11,29 @@ jmh {
 
     // Force a System.gc() between iterations so heap state is reset before each
     // measurement starts. Without this, a GC pause can land inside the measurement
-    // window and show up as a spurious regression in CI. Safe under G1 (JDK 21
-    // default) — avoid combining with -XX:+DisableExplicitGC, which would make this
-    // a silent no-op.
+    // window and show up as a spurious regression in CI.
     forceGC = true
 
     // -Xbatch forces JIT compilation to run synchronously on the benchmark thread
     // instead of on a background compiler thread. With background compilation, the
     // C2 thread can kick in mid-measurement and steal CPU cycles, producing jitter
     // that shows up as a regression on unchanged code. Synchronous compilation
     // front-loads the cost into warmup where it belongs.
-    jvmArgs.set(listOf("-Xbatch"))
+    //
+    // -Xms/-Xmx: Fixed heap size eliminates resize pauses and adaptive sizing noise.
+    // -XX:+AlwaysPreTouch: Pre-touches all heap pages at startup so lazy page faults
+    //   don't land inside measurement windows.
+    // -XX:+UseSerialGC: Single-threaded GC is the most predictable — no concurrent
+    //   GC threads, no G1 refinement jitter.
+    // -XX:-UseAdaptiveSizePolicy: Prevents the GC from dynamically resizing
+    //   young/old generations between runs.
+    jvmArgs.set(listOf(
+        "-Xbatch",
+        "-Xms2g", "-Xmx2g",
+        "-XX:+AlwaysPreTouch",
+        "-XX:+UseSerialGC",
+        "-XX:-UseAdaptiveSizePolicy",
+    ))
 }
 
 sourceSets {