Skip to content

Latest commit

 

History

History
212 lines (169 loc) · 11.3 KB

README.md

File metadata and controls

212 lines (169 loc) · 11.3 KB
Raw

ion-java Benchmarking CLI

This tool allows users to...

  • Determine which ion-java configurations perform best
  • Compare ion-java to Java implementations of other serialization formats (not yet implemented)

...for the individual users' data and access patterns.

Additionally, allows ion-java developers to...

  • Determine the impact of a proposed change
  • Decide where investments should be made in improving performance

...by generating results from a variety of real-world data and access patterns.

The tool uses the JMH microbenchmarking library and produces results in formats generated by JMH.

The following stats will be included in benchmark results:

  • Speed (or throughput)
  • Heap usage
  • Garbage collection statistics (number of GCs, total time taken, average size of various GC generations)
  • Size of the data (the input data for read benchmarks and the output data for write benchmarks)

Installation

Building the tool generates a self-contained executable jar.

To build the tool with the latest version of the ion-java library, simply run

mvn clean install

from the directory containing pom.xml. When the build completes, a jar named ion-java-benchmark-cli-<version>-jar-with-dependencies.jar will be present in the target/ subdirectory.

This jar may be copied to any location and may be executed using java -jar <path-to-jar>. Users may find it convenient to create an alias to java -jar <path-to-jar> named, e.g., ion-java-benchmark to simplify invocation of the tool.

Note: the Maven build will look for the latest ion-java version available in the local Maven repository. Users may use the tool with a custom or in-development ion-java version by installing the desired version into the local Maven repository. To determine which ion-java version the tool is using, use the --version command. After switching ion-java versions, the executable jar must be rebuilt.

Examples

(Note: these examples assume the invocation of the tool has been aliased to ion-java-benchmark.)

Benchmark a full-traversal read of example.10n from file using the IonReader API, with 10 warmups, 10 iterations, and 1 fork, printing the results to stdout in JMH’s standard text format.

ion-java-benchmark read example.10n

Benchmark a fully-buffered write of binary Ion data equivalent to example.10n to file using the IonWriter API, with 10 warmups, 10 iterations, and 1 fork, printing the results to stdout in JMH’s standard text format.

ion-java-benchmark write example.10n

Benchmark a write of binary Ion data equivalent to the first 1,000 top-level values in example.10n to in-memory bytes using the IonWriter API, flushing after every 100 top-level values. Produce results for both 0-byte length preallocation and 2-byte length preallocation to facilitate comparison of both settings.

ion-java-benchmark write --io-type buffer \
                         --limit 1000 \
                         --ion-flush-period 100 \
                         --ion-length-preallocation 0 \
                         --ion-length-preallocation 2 \
                         example.10n

Profile a sparse read of example.10n from file, materializing only the values that match the paths specified in paths.ion, using ion-java-path-extraction. This process will repetitively execute until manually terminated, allowing the user to attach a tool for gathering performance profiles.

ion-java-benchmark read --profile --paths paths.ion example.10n

Benchmark a fully-buffered write of binary Ion data equivalent to example.10n both with and without using shared symbol tables. The file tables.ion contains a sequence of Ion symbol tables.

ion-java-benchmark write --ion-imports-for-benchmark tables.ion \
                         --ion-imports-for-benchmark none \
                         example.10n

Benchmark a full-traversal read of data equivalent to exampleWithImports.10n, which declares the shared symbol table imports provided by inputTables.ion, re-encoded (if necessary) using the shared symbol tables provided by benchmarkTables.ion, inputTables.ion, and no shared symbol tables. Produce results from using both the DOM and IonReader APIs.

ion-java-benchmark read --ion-imports-for-input inputTables.ion \
                        --ion-imports-for-benchmark benchmarkTables.ion \
                        --ion-imports-for-benchmark auto \
                        --ion-imports-for-benchmark none \
                        --ion-api dom \
                        --ion-api streaming \
                        exampleWithImports.10n

Tips

As the JMH output warns: "Do not assume the numbers tell you what you want them to tell." Benchmarking on the JVM is hard. There is non-deterministic behavior that can lead to high variance between iterations. Be suspicious of benchmark results with a reported Error that is a high percentage of the Score. Aim for an Error percentage of less than 10%.

To reduce Error, try increasing the number of warmup iterations, timed iterations, and forks. To ensure the JVM is properly warmed up, benchmarks should include enough warmup iterations to allow for the scores to stabilize. This often takes at least 20 seconds. Benchmarks should be run on idle systems. Background processes competing for resources can lead to higher variance, especially for benchmarks with a short execution time per invocation.

The default benchmark mode is SingleShotTime, meaning that the reported score is the result of a single invocation of the benchmark method. This works well for medium and large input data that takes on the order of seconds per invocation, but leads to higher variance for input data that takes only milliseconds or microseconds per invocation. Such data should be used with one of the other modes, each of which generates an iteration score by averaging the score of multiple invocations of the benchmark method. For very small data, it may also be necessary to change the reported time unit to provide enough granularity to observe differences between trials.

Both the read and write benchmark commands involve a setup phase that occurs before the benchmark begins. However, due to a quirk in the JMH implementation, this phase occurs after JMH prints Warmup Iteration 1: to the output. This can make it seem like the first warmup iteration takes an excessive amount of time or is deadlocked, but give it a chance to complete. Once the scores stabilize, if the iterations take longer than you're willing to wait, consider using the --limit option to limit the amount of data processed by the benchmark. For write benchmarks, using --limit may be necessary depending on the size of the input data and the memory constraints of the system, as the setup phase involves generating write instruction lambdas and storing them in memory.

For developers

Adding an option

Adding an option involves the following steps:

  1. In Main, add the option to the USAGE and OPTIONS strings, mimicking the existing format.
  2. If the option applies to both the read and write commands, add parameterization logic to the OptionsMatrixBase constructor using the existing helper methods. If the option applies only to the read or the write command, add this logic to the ReadOptionsMatrix or WriteOptionsMatrix constructor, respectively. These classes are responsible for generating the complete set of combinations for the chosen set of option values, which corresponds to the complete set of benchmark trials to be run.
  3. If the option applies to both the read and write commands, add parsing logic to the OptionsCombinationBase constructor using the existing helper methods. If the option applies only to the read or the write command, add this logic to the ReadOptionsCombination or WriteOptionsCombination constructor, respectively. These classes are responsible for representing a single combination of option values, which corresponds to a single benchmark trial.
  4. Determine where to place the logic that uses the new option. If the option applies to all formats or affects how a resource is constructed or configured, it may make sense to add a factory method to OptionsCombinationBase, ReadOptionsCombination, or WriteOptionsCombination and invoke this method in any MeasurableTask implementations to which the option applies. As an example, see OptionsCombinationBase.newInputStream. If the option applies to all formats but only the write or read command, then using the option within MeasurableReadTask or MeasurableWriteTask may make sense. As an example, see MeasurableWriteTask.getTask. If the option applies to multiple formats, but not all formats, it may make sense to add a utility function or use inheritance. As an example, see IonUtilities, which is used by IonMeasurableReadTask and IonMeasurableWriteTask. If the option is limited to use with a particular format and command, then its logic may belong in the concrete MeasurableTask implementation for that format/command combination. As an example, the --ion-reader option only applies to the read command when used with either Ion text or binary, so the logic that uses the option is contained within IonMeasurableReadTask.
  5. In OptionsTest, add tests that exercise all values for this option (if enumerated) or a variety of values, for all commands to which it applies.

Commit d37b6fd demonstrates the steps required to add a new option.

Adding a format

Adding support for an additional serialization format involves the following steps.

  1. Identify the Java library (or libraries) that provides the reader/writer implementation for that format in Java. Add an open-ended dependency on that library to pom.xml.
  2. Add a value to the Format enum to represent the new format.
  3. In Main's OPTIONS string, edit the entry for the --format command to allow for the new format.
  4. Using IonMeasurableReadTask and IonMeasurableWriteTask as examples, create concrete implementations of MeasurableReadTask and MeasurableWriteTask for the new format.
  5. Implement the inherited abstract methods in the new Format enum value. This involves adding logic to convert between formats and to instantiate the MeasurableTask classes created in the previous step. The convert implementations for the existing Format values will need to be updated as well to support conversions from the new format. Add logic to Format.classify to determine whether a file contains data in the new format.
  6. Follow the steps from the Adding an option section above to add any format- or library-specific options to the CLI.
  7. In OptionsTest, add tests that thoroughly exercise the new format. Add data in the new format into the test directory for use in tests.
  8. Build the tool using mvn clean install. Using various samples of data in the new format, run the tool by hand to make sure everything looks correct and the benchmark results look reasonable.
  9. Add at least one example of using the new format to the EXAMPLES string in Main. Copy this example into the Examples section of this README.

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.