CrossVersionPerformanceTestRunnerTest.groovy

Instead of an arbitrary threshold we now use the Mann-Whitney U test

to calculate how confident we can be in the results.

For now we'll require a 99% confidence before reporting a regression.

We can tweak that value based on whether we miss too many

regressions or report too many non-regressions.

Previously we were using a workaround where odd runs were removed from measurements, and executed a `clean` build instead of the measured build.

Performance tests can now specify `cleanTasks`, similar to how they specify `tasksToRun`. These `cleanTasks` will be executed before each run (warm-up and measurement runs alike).

A new column is added to performance test tables to track this new information. It is a nullable column to allow for test results added by older versions of Gradle.

I've updated the task output caching tests and the Maven vs. Gradle comparisons to declare `cleanTasks` instead of the old hack with the odd-even runs.

The many measurements that we injected into the build under test were

skewing our measurements to the point of making them unreliable or

unrealistic. We now only measure end to end build time. Here's a breakdown

with the rationale for removing each other measurement:

- configuration time: can be done by having a `gradle help` scenario instead

- execution time: the user does not care whether a long build is stuck in execution or configuration

- setup/teardown: was ill-defined anyway, basically total - configuration - execution

- JIT compile time: this is nothing we can influence and thus pointless to measure

- Memory usage: Was only measured at one point in the build, which doesn't tell us anything about

any problems at any other point in the build

- GC CPU time: If this increases we'd see it in total execution time

Generally, looking at the graphs has never pointed us directly at the problem, we always need to

profile anyway. So instead of skewing our measurements with lots of profling code, we should

instead use a dedicated profiling job to measure if we actually see a regression.

Memory usage can be tested indirectly by giving each scenario a reasonable amount of memory.

If memory usage rises above that reasonable limit, we'd see execution time rise, telling us about

the regression. Generally, we do not optimize for smallest memory usage, but for fastest execution

with reasonable memory overhead.

This change also removes all JVM tweaking and wait periods which we introduced in an attempt to

make tests more predictable and stable. These tweaks have not really helped us achieve more stable

tests and have often done the opposite. They also add lots of complexity and make our tests more

unrealistic. A real users will not add all these JVM options to Gradle.

Note that versionBase from the build-receipt.properties is

only used in `incoming-distributions.gradle`.

PR #798

- optimize for stable results

- enables using performance test specific build context when an instance

is properly wired

This commit improves the integration of Honest Profiler by making it possible to post-process the log files and generate a flame graph.

For this to be possible, the `org.gradle.performance.honestprofiler` system property must be set to a directory where flame graphs will

be exported.

It is expected to find 2 environment variables:

- `HP_HOME_DIR` must point to a valid Honest Profiler installation (https://github.com/RichardWarburton/honest-profiler)

- `FG_HOME_DIR` must point to a valid FlameGraph installation (https://github.com/brendangregg/FlameGraph)

In case those are not set, the collector will try to find them in `~/tools/honest-profiler` and `~/tools/FlameGraph` respectively.

The graphs are *not* integrated into the performance reports yet.

- it's faster to do profiling when running only the current version

- There was previously a concern that this would break the

BuildScansPerformanceTest, but that test doesn't actually use the

BaselineVersion class at all.

This reverts commit 784d7476556e63c8d7f15919ae9875c3b7181aa3.

This commit introduces a new flag (`--checks`) or system property (`org.gradle.performance.execution.checks`)

that can be used to control which regressions are going to be looked for in a performance test. By default,

we're looking for regressions in memory and speed. With this flag it is possible to control precisely what

to look for:

- `all` is the default and will look for both memory and speed regressions

- `memory` will only check for memory regressions

- `speed` will only check for speed regressions

- `none` will not check for regressions

The `memory` and `speed` values are useful when bisecting a regression: they let us ignore potential statistical

artifacts of measurement of a category we're not interested in.

The `none` value is interesting for historical builds, where we want to collect all measurements, but not

fail the build if a regression is encountered.