Working with JMH (Microbenchmarking)

1. Setting Up JMH Project

Step	Detail
Generate	`mvn archetype:generate -Dgroup=org.openjdk.jmh -DartifactId=jmh-archetype`
Dependencies	jmh-core, jmh-generator-annprocess
Build	Shade JMH runner JAR
Run	`java -jar target/benchmarks.jar`

2. Creating Benchmark Methods (@Benchmark)

Example: Simple benchmark

@Benchmark
public int hashCode(MyState s) {
    return s.value.hashCode();
}

Rule	Detail
Return result	Or use Blackhole
No I/O	Unless intentionally measuring it

3. Using Benchmark Modes

Mode	Reports
Throughput	ops/time
AverageTime	time/op
SampleTime	Sampled time distribution
SingleShotTime	One iteration (cold)
All	All of the above

4. Configuring Warmup (@Warmup)

Param	Default
iterations	5
time	10 s
batchSize	1
Purpose	Reach steady state (JIT, caches)

5. Configuring Measurement (@Measurement)

Param	Use
iterations	5–10 typical
time	10 s typical
forks	JVM forks (default 5)
threads	Concurrent worker count

6. Using State Objects (@State)

Example: State

@State(Scope.Benchmark)
public static class MyState {
    String value;
    @Setup public void init() { value = "hello"; }
}

Aspect	Detail
Lifecycle	Linked to Scope
Injection	Pass as @Benchmark method param

7. Understanding Scope

Scope	Sharing
Benchmark	One instance for all threads
Group	Shared in @Group methods
Thread	Per-thread instance

8. Using Setup and TearDown (@Setup, @TearDown)

Level	Runs
Trial	Once per benchmark run
Iteration	Each iteration
Invocation	Every @Benchmark call (use sparingly — bias risk)

9. Avoiding Dead Code Elimination (Blackhole)

Example: Blackhole

@Benchmark
public void multi(Blackhole bh) {
    bh.consume(work1());
    bh.consume(work2());
}

Issue	Fix
Unused result	Return it OR Blackhole.consume()
Constant folding	Read input from @State

10. Running Benchmarks

Command	Use
`java -jar benchmarks.jar`	Run all
`java -jar benchmarks.jar Pattern`	Filter by name regex
`-prof gc`	Allocation profiler
`-prof async:libPath=...;output=flamegraph`	async-profiler
`-rf json -rff out.json`	JSON output

11. Analyzing Results

Field	Meaning
Score	Mean throughput / time
Error	99.9% confidence half-width
Ratio	Compare to baseline
Visualize	jmh-visualizer (web UI)

12. Understanding JMH Pitfalls

Warning: Microbenchmarks are easy to write but hard to interpret. Always: (1) cross-check with macro-benchmark, (2) verify dead-code isn't elided, (3) use multiple forks, (4) account for warmup, (5) compare on the same JVM/hardware.

Pitfall	Effect
Constant folding	JIT precomputes result
Inlining hides cost	Use Blackhole
No warmup	Measures interpreter
Single fork	Class init bias