NewRegressionFramework

We'd like to revamp the regression tests by moving to a new framework. This page is intended to host a discussion of features and design for the new framework.

Ali's plan for a new implementation

• Use pytest
• It has, by far, the best documentation of any of the python testing frameworks and seems to be the most active
• Seems to be completely extensible via python plugins and hooks
• Supports outputting JUnit XML incase we want to use a continuous integration solution such as Jenkins or Hudson
• The pytest xdist and plugin support running tests on multiple-cpus or multiple machines
• Good collection of tasks and tutorials

How things would work

• Marks may be assigned to tests either with python decorators or a class attribute if we want to stay python 2.5 compatible
  • The decorators would probably include cpu model, memory system, ISA, mode, and run length.
  • We might want to use pytest_addoption to be able to pass lists specifically for each of the decorators and generate tests that match appropriately with this
  • Alternatively we could use pytest-markfiltration although the syntax can be rather contrived

Outstanding Questions

• How would we do test discovery?
  • pytest will search py files looking for tests
  • Files can match a pattern, classes in files can match a pattern or functions can match a pattern
  • or it can only match things that inherit from Python UnitTest
• Should we use xunit style or func args style setups?
• Should we have a class that inherits from Python.UnitTest and does the heavy lifting or should we have a completely separate class that does the heavy lifting and use a factory class to create a bunch of instances of the seperate class?
• Should gem5 be called as a library or on the command line?
• How should we store output files? Same way we do now? should each directory just have a __init__.py and then the tests can be referred to as long.linux_boot.arm.linux.o3?

Desirable features

• Ability to add regressions via EXTRAS
  • For example, move eio tests into eio module so we don't try to run them when it's not compiled in
• Ability to not run regressions for which binaries or other inputs aren't available
  • With maybe some nice semi-automated way of downloading binaries when they're publicly available
• Better categorization of tests, and ability to run tests by category, e.g.:
  • by CPU model
  • by ISA
  • by Ruby protocol
  • by length
• More directed tests that cover specific functionality and complete faster. Running spec benchmarks is important but spends a lot of time doing the same thing over and over. Those should only be a component of our testing, not almost all of it like it is now. This is a desirable feature of our testing strategy, not necessarily something that impacts the regression framework.
• Better checkpoint testing
  • some of this doesn't really depend on the regression framework, just needs new tests
  • e.g., integrating util/checkpoint-tester.py
• Support for random testing (e.g., for background testing processes)
  • Random latencies?
  • Random testing a la memory testers but with different seeds, longer intervals
• Decouple from SCons
  • Avoid having scons dependency bugs force unnecessary re-running of tests, particularly for update-refs
  • Don't rely on scons to run jobs... running scons -j8 with a bunch of tests and a batch queing system means that 8 cpus are consumed, even if there is only one job running.
  • Either make scons be able to submit the jobs or have something else that manages the jobs and their completion status
• Easy support for running separate tests where only the input parameters differ
  • For example, several protocols utilize different state transitions depending on configuration flags. It would be great if we could test these without having to create new directories and tests.
  • Similarly, we could/should test topologies this way as well.
• Automated way to use nightly regressions as a basis for updating "m5-stable"
  • How do you identify the last working revision? (from Ali)
  • Maybe need a bug-tracking system so we could record facts like "changeset Y fixes a bug introduced in changeset X" then we could automatically exclude changesets between X and Y, but we don't have that. (from stever)
• Better definitions of success criteria.
  • E.g. Stats were changed, but output is all still correct vs simply passed and failed. (Passed, stats diffs, failed)
  • For example you could say that the terminal output changing is fail, or the stdout and spec binary outputs changing are failed, but a 1% difference in stats is a stats difference, which needs to be addresses
  • I envision this as providing reasonable certainty that if you create a change you know will modify the stats, you have a quick verification that nothing broke horribly before updating the stats.

Implementation ideas

Just ideas... no definitive decisions have been made yet.

• Use Python's unittest module, or something that extends it such as nose
• Use SCons to manage dependencies between binaries/test inputs and test results, but in a different SCons invocation (i.e., in its own SConstruct/SConscript)