Producing Fuchsia performance test results

This page lists libraries for producing Fuchsia performance test results files.

The format used for Fuchsia performance test results is the fuchsiaperf.json format. All performance tests that run on Fuchsia Infra and whose results are tracked by the Fuchsia tools produce results in the fuchsiaperf.json format. All of these tests are run through the Python-based Lacewing framework or the Dart-based SL4F testing framework, though in many cases the Python code is hidden from view and developers just need to write GN.

Options

There are multiple options for generating fuchsiaperf.json files, depending on which programming language you want to use.

The low level options listed below are thin wrappers for outputting fuchsiaperf JSON files, whereas the higher level options make more assumptions about the type of performance test being written.

• Python:

  • High level: You can use the Python trace_processing library to extract performance metrics from Fuchsia traces. This approach is useful if you have an existing correctness test and you want to extend it to also produce performance results. In that case, it is common to modify the software-under-test to generate extra trace events.

    An example is perftest_trace_events_test which uses the trace_processing library to extract a set of events.

    Another example is flatland_benchmark which uses the trace_processing.app_render library to extract a set of events, then uses perf_publish to upload the result.

    The test should use the python_perf_test template as it includes all the necessary dependencies for trace processing and metrics publishing.

• Dart (Deprecated):

  • High level: You can use the Dart trace_processing library to extract performance metrics from Fuchsia traces. This approach is useful if you have an existing correctness test and you want to extend it to also produce performance results. In that case, it is common to modify the software-under-test to generate extra trace events.

  • Low level: You can use the TestCaseResults class to generate entries for fuchsiaperf.json files. This is commonly used with the trace_processing library, but it can also be used separately.

  • High or low level: From Dart, you can run a subprocess that generates a fuchsiaperf.json file. The subprocess can run code written in a language other than Dart. There are various Dart SL4F wrappers that do this.

• C++: The perftest C++ library provides two interfaces:

  • High level: You can use perftest.h to create microbenchmarks. In this context, a microbenchmark is a test where we run an operation repeatedly, in isolation, and measure its running time. New microbenchmarks can be added to src/tests/microbenchmarks/, or they can be added elsewhere in the source tree if they are significantly different from the tests in that directory.

  • Low level: You can use perftest/results.h to generate fuchsiaperf.json files more directly.

• Rust:

  • High level: You can use the fuchsia-criterion Rust library to create microbenchmarks.

  • Low level: You can use the Fuchsiaperf Rust library to generate fuchsiaperf.json files.

• Go:

  • Low level: You can use the go-benchmarking library to generate fuchsiaperf.json files.

Declaring tests in GN

For a wide amount of benchmarks, we only need to write the target side component, run it, process the fuchsiaperf file outputted by the test and publish the results. The template fuchsia_component_perf_test simplifies this. The template python_perf_test can be used by other host-side tests written in Python that deal with performance metrics.