-
Notifications
You must be signed in to change notification settings - Fork 38.2k
Micro Benchmarks
This document describes the strategy for performance micro-benchmarks in Spring Framework.
Micro-benchmarks are designed for optimizing code in the Framework, when hot code paths and usage profiles are known. This approach does not replace profilers and full performance benchmarks (complete applications, with I/O and network latency).
Spring Framework uses the JMH harness for building and running micro-benchmarks.
JMH is integrated with the Spring Framework build thanks to the JMH Gradle plugin.
This plugin is applied to all spring-* modules and creates a jmh configuration for them.
Benchmark sources are located in src/jmh/java in each Spring Framework module.
The Gradle plugin is applied and configured in /gradle/spring-module.gradle.
JMH helps you to write and run benchmarks while avoiding as much as possible common benchmark pitfalls. This is why JMH looks at the benchmark sources and generates classes that will run the benchmark.
The Gradle plugin provides tasks for building the JMH jars (./gradlew jmhJar) and running the benchmarks (./gradlew jmh).
The next section will explain how to use existing benchmarks in the Spring Framework codebase for optimization work.
While ./gradlew jmh will run benchmarks from the CLI, this command will run all available benchmarks in the project,
unless you modify the existing plugin configuration on a per-usage basis.
The easiest way to work with an existing benchmark is the following.
First, we need to build the JMH jar for the spring module we’re working with, e.g. ./gradlew :spring-core:jmhJar.
We can browse the source repository or list the available benchmarks on the CLI and find the one(s) that we’d like to work with,
java -jar spring-core/build/libs/spring-core-5.3.0-SNAPSHOT-jmh.jar -l.
Each benchmark can provide benchmark configuration with annotations (benchmark type, JVM configuration, parameters). We need to look at the benchmark class(es)/methods and familiarize ourselves with them; this step is required to figure out which benchmark options we want to enable.
For example, we can run the MimeTypeBenchmark.cachedParserSomeMimeTypes benchmark, with 2 JVM forks, the GC profiler, and specific benchmark parameters with:
java -jar spring-core/build/libs/spring-core-5.3.0-SNAPSHOT-jmh.jar -f1 -prof gc -p customTypesCount=10,20 MimeTypeBenchmark.cachedParserSomeMimeTypesThe benchmark options are really important, as the specific parameters are part of the benchmark design;
JMH parameters can help you test a wide variety of runtime/workload profiles (like concurrency).
All JMH options can be listed on the CLI using java -jar spring-core/build/libs/spring-core-5.3.0-SNAPSHOT-jmh.jar -h.
We can then change the source code, rebuild the JMH jar, re-run and compare results.
Note that it’s not always easy to compare two implementations and many aspects should be taken into account:
-
differents JVM version and options
-
the error magin of the results
-
number of iterations, forks and worker threads for the benchmark run (JMH CLI options)
-
benchmark parameter values for the specific benchmark
-
running other benchmarks on the same code, as maybe a specific optimization made another use case worse
In any case, questioning the benchmark design and improving it is always useful - see next section!
With the current build infrastructure, adding a new Benchmark can be done easily: adding a new class under src/jmh/java is enough.
For consistency, all benchmark classes should be named *Benchmark.
If additional dependencies are required for the benchmarks, they can be added to the jmh configuration in the dependencies section of the Spring module:
dependencies {
jmh 'org.example:sample:1.0.0'
}Designing a micro-benchmark can be hard, and JMH won’t automatically prevent all benchmarking pitfalls from happening. The JVM, OSes and CPUs have many optimization strategies - JMH provides tools to avoid those, but good benchmark design is still essential.
JMH provides many benchmark samples to learn about design.
We’ll use here a sample benchmark and discuss best practices. Let’s say we’d like to measure the performance of our MimeTypeUtils MIME type parser.
Because this code is executed many times at runtime, we’ll design the benchmark to measure and optimize throughput.
JMH supports many benchmark modes.
A basic benchmark can use the JMH annotation infrastructure: We’d like to take a set of various raw MIME types and parse them with the parser.
@BenchmarkMode(Mode.Throughput)
public class MimeTypeBenchmark {
@Benchmark
public void parseMimeTypes() {
final List<String> mimeTypes = Arrays.asList("application/json","text/html", ...);
for(String type : mimeTypes) {
MimeType parsed = MimeTypeUtils.parseMimeType(type);
}
}
}There are many issues with this benchmark and we’ll try to fix them. First, our input is predictable and Constant Folding is likely to happen. We can extract this data into a shared state.
@BenchmarkMode(Mode.Throughput)
public class MimeTypeBenchmark {
@State(Scope.Benchmark)
public static class BenchmarkData {
public List<String> mimeTypes;
public BenchmarkData() {
this.mimeTypes = Arrays.asList("application/json","text/html", ...);
}
}
@Benchmark
public void parseMimeTypes(BenchmarkData data) {
for(String type : data.mimeTypes) {
MimeType parsed = MimeTypeUtils.parseMimeType(type);
}
}
}Our benchmark still has a problem of Dead Code Elimination; usually JMH takes care of this if the operation result is returned by the method, but in this case we’ve got multiple results and no easy way to merge them. We’re going to use Blackholes to solve that problem.
@BenchmarkMode(Mode.Throughput)
public class MimeTypeBenchmark {
@State(Scope.Benchmark)
public static class BenchmarkData {
public List<String> mimeTypes;
public BenchmarkData() {
this.mimeTypes = Arrays.asList("application/json","text/html", ...);
}
}
@Benchmark
public void parseMimeTypes(BenchmarkData data, Blackhole, bh) {
for(String type : data.mimeTypes) {
bh.consume(MimeTypeUtils.parseMimeType(type));
}
}
}Even with this new version, we need to carefully consider using loops in benchmark methods.
As an improvement, we could consider adding @Param annotations to parameterize the size of the raw MIME types to parse.
We can also create a baseline method to compare with/without the parsing execution:
@BenchmarkMode(Mode.Throughput)
public class MimeTypeBenchmark {
@State(Scope.Benchmark)
public static class BenchmarkData {
public List<String> mimeTypes;
public BenchmarkData() {
this.mimeTypes = Arrays.asList("application/json","text/html", ...);
}
}
@Benchmark
public void baseline(BenchmarkData data, Blackhole, bh) {
for(String type : data.mimeTypes) {
bh.consume(type);
}
}
@Benchmark
public void parseMimeTypes(BenchmarkData data, Blackhole, bh) {
for(String type : data.mimeTypes) {
bh.consume(MimeTypeUtils.parseMimeType(type));
}
}
}In all cases, it’s a good idea to try different approaches, check out data provided by JMH profilers, and discuss the benchmark with colleagues!
This sample benchmark is quite simple, and other problem spaces (like controlling concurrency) can be dealt with by using other features showcased in the samples.