JMH is the state-of-the-art Java Microbenchmarking Harness. Setting it up (especially with Gradle) and learning how to use it can be a bit difficult; hopefully this repo makes this process easier for others.
# Runs both 'clean' and 'shadowJar' by default, which is what you want
./gradlew
# Runs the benchmark that matches the provided regex; pass -h instead of a regex
# to see all JMH options.
java -jar build/libs/benchmarks.jar "HelloWorld"
# Ex: 1 run (fork), 8 warmup iterations, 10 measurement iterations
java -jar build/libs/benchmarks.jar "HelloWorld" -f 1 -wi 8 -i 10
Does NOT use the jmh-gradle-plugin which is confusing, brittle and difficult to use correctly.
The absolute best way to learn how to use JMH is to read through the official JMH samples. All of them are included in this repo. Start by reading the first sample file (they are all numbered) and proceed from there.
READ ALL THE JMH SAMPLES before creating any benchmarks! There are many, many pitfalls that need to be avoided, lest your benchmarks end up producing false results. The JMH samples list out issues to watch out for and how to avoid them. Here's a summary:
- Unintentional dead code elimination (JMH sample 8).
- Computations getting constant-folded (JMH sample 10).
- Loops in benchmarks having iterations merged by JIT (JMH sample 11).
- JMH sample 34 covers the same issue and provides good advice.
- False Sharing AKA independent fields on the same cache line affecting each other (JMH sample 22).
- Test data helping or hurting branch prediction (JMH sample 36).
- Differences in cache access (JMH sample 37).
- Incorrect per-invocation benchmark setup (JMH sample 38).
- You need warmup iterations because of JVM and JIT warmup. How many depends on the benchmark, but probably no less than 5. A safer number is 10.
- The more measurement iterations you use, the smaller the error margin reported by JMH at the end! A solid choice is 20 iterations.
- If you really want to get that measurement error margin to be as small as possible, reboot your machine and run your JMH benchmarks without anything else running on your system. No browser, IDE etc taking up CPU time. You can go even further by turning off dynamic CPU frequency scaling in your system BIOS.
- Sometimes you need to know more than just the average time to run your
benchmark.
@BenchmarkMode(Mode.SampleTime)can show you a distribution with percentiles for the time it takes to run your bench method. - Be extremely wary of dead code elimination; see JMH sample 8. You should basically always be returning a value from your benchmark or using a Blackhole (see JMH sample 9).
- To test performance of code where several threads are doing different work (e.g. one thread reading data, one writing data), use JMH thread groups. See JMH sample 15 for details.
- If you need to use some custom operation/event counters in your
benchmarks, use JMH's
AuxCounters. See JMH sample 23. - If you really need a specific invocation count to avoid variance,
instead of looping in the benchmark, use JMH
batchSizeto control the number of calls per invocation. See JMH sample 26. - Use JMH
@Paramto control benchmark configuration. For instance, seeing performance change as the size of an array changes. See JMH sample 27. - JMH also has some built-in profilers you can use. See JMH sample
35. NOTE: Using multiple benchmark forks is even more important if
using profilers to reduce measurement error margin.
- The
stackprofiler is a simple sampling profiler that can show hot methods. - The
gcprofiler is amazing for understanding garbage creation rate etc. - The
perfnormprofiler uses Linuxperfcommand to read CPU hardware counters for branch misses, cache loads/stores/misses etc. It also normalizes the values to benchmark iterations.
- The
Blackhole.consumeCPU(tokens)is the JMH-recommended way of ensuring the CPU does something (in case you need that). Useful for benchmarks where you use the GC profiler to see garbage creation rate and don't actually care about the measurement time.
The full output is below; the "interesting" commands are:
-f <int>for number of runs-wi <int>for number of warmup iterations within a run-i <int>for number of measurement iterations within a run
Note that all of these can be changed with JMH annotations on the benchmark itself (command line options override).
Usage: java -jar ... [regexp*] [options]
[opt] means optional argument.
<opt> means required argument.
"+" means comma-separated list of values.
"time" arguments accept time suffixes, like "100ms".
Command line options usually take precedence over annotations.
[arguments] Benchmarks to run (regexp+). (default: .*)
-bm <mode> Benchmark mode. Available modes are: [Throughput/thrpt,
AverageTime/avgt, SampleTime/sample, SingleShotTime/ss,
All/all]. (default: Throughput)
-bs <int> Batch size: number of benchmark method calls per
operation. Some benchmark modes may ignore this
setting, please check this separately. (default:
1)
-e <regexp+> Benchmarks to exclude from the run.
-f <int> How many times to fork a single benchmark. Use 0 to
disable forking altogether. Warning: disabling
forking may have detrimental impact on benchmark
and infrastructure reliability, you might want
to use different warmup mode instead. (default:
10)
-foe <bool> Should JMH fail immediately if any benchmark had
experienced an unrecoverable error? This helps
to make quick sanity tests for benchmark suites,
as well as make the automated runs with checking error
codes. (default: false)
-gc <bool> Should JMH force GC between iterations? Forcing
the GC may help to lower the noise in GC-heavy benchmarks,
at the expense of jeopardizing GC ergonomics decisions.
Use with care. (default: false)
-h Display help, and exit.
-i <int> Number of measurement iterations to do. Measurement
iterations are counted towards the benchmark score.
(default: 1 for SingleShotTime, and 20 for all other
modes)
-jvm <string> Use given JVM for runs. This option only affects forked
runs.
-jvmArgs <string> Use given JVM arguments. Most options are inherited
from the host VM options, but in some cases you want
to pass the options only to a forked VM. Either single
space-separated option line, or multiple options
are accepted. This option only affects forked runs.
-jvmArgsAppend <string> Same as jvmArgs, but append these options after the
already given JVM args.
-jvmArgsPrepend <string> Same as jvmArgs, but prepend these options before
the already given JVM arg.
-l List the benchmarks that match a filter, and exit.
-lp List the benchmarks that match a filter, along with
parameters, and exit.
-lprof List profilers, and exit.
-lrf List machine-readable result formats, and exit.
-o <filename> Redirect human-readable output to a given file.
-opi <int> Override operations per invocation, see @OperationsPerInvocation
Javadoc for details. (default: 1)
-p <param={v,}*> Benchmark parameters. This option is expected to
be used once per parameter. Parameter name and parameter
values should be separated with equals sign. Parameter
values should be separated with commas.
-prof <profiler> Use profilers to collect additional benchmark data.
Some profilers are not available on all JVMs and/or
all OSes. Please see the list of available profilers
with -lprof.
-r <time> Minimum time to spend at each measurement iteration.
Benchmarks may generally run longer than iteration
duration. (default: 1 s)
-rf <type> Format type for machine-readable results. These
results are written to a separate file (see -rff).
See the list of available result formats with -lrf.
(default: CSV)
-rff <filename> Write machine-readable results to a given file.
The file format is controlled by -rf option. Please
see the list of result formats for available formats.
(default: jmh-result.<result-format>)
-si <bool> Should JMH synchronize iterations? This would significantly
lower the noise in multithreaded tests, by making
sure the measured part happens only when all workers
are running. (default: true)
-t <int> Number of worker threads to run with. 'max' means
the maximum number of hardware threads available
on the machine, figured out by JMH itself. (default:
1)
-tg <int+> Override thread group distribution for asymmetric
benchmarks. This option expects a comma-separated
list of thread counts within the group. See @Group/@GroupThreads
Javadoc for more information.
-to <time> Timeout for benchmark iteration. After reaching
this timeout, JMH will try to interrupt the running
tasks. Non-cooperating benchmarks may ignore this
timeout. (default: 10 min)
-tu <TU> Override time unit in benchmark results. Available
time units are: [m, s, ms, us, ns]. (default: SECONDS)
-v <mode> Verbosity mode. Available modes are: [SILENT, NORMAL,
EXTRA]. (default: NORMAL)
-w <time> Minimum time to spend at each warmup iteration. Benchmarks
may generally run longer than iteration duration.
(default: 1 s)
-wbs <int> Warmup batch size: number of benchmark method calls
per operation. Some benchmark modes may ignore this
setting. (default: 1)
-wf <int> How many warmup forks to make for a single benchmark.
All iterations within the warmup fork are not counted
towards the benchmark score. Use 0 to disable warmup
forks. (default: 0)
-wi <int> Number of warmup iterations to do. Warmup iterations
are not counted towards the benchmark score. (default:
0 for SingleShotTime, and 20 for all other modes)
-wm <mode> Warmup mode for warming up selected benchmarks.
Warmup modes are: INDI = Warmup each benchmark individually,
then measure it. BULK = Warmup all benchmarks first,
then do all the measurements. BULK_INDI = Warmup
all benchmarks first, then re-warmup each benchmark
individually, then measure it. (default: INDI)
-wmb <regexp+> Warmup benchmarks to include in the run in addition
to already selected by the primary filters. Harness
will not measure these benchmarks, but only use them
for the warmup.
JMH samples (which are included in the source code) are licensed under the 3-clause BSD license.
The rest of the code in this repo is licensed under Apache v2.