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
  • If we used py.test on the build directory that would find the eio code or whatever else is in there and pick them up. This could be through any of the methods described above --Saidi 18:09, 22 August 2011 (PDT)
• Ability to not run regressions for which binaries or other inputs aren't available
  • This could be done easily based on the BaseClass that we use and tell about the file dependencies. e.g. if not dependencies: pytest.skip("test binaries unavailable") --Saidi 18:09, 22 August 2011 (PDT)
  • With maybe some nice semi-automated way of downloading binaries when they're publicly available
  • This might be harder --Saidi 18:09, 22 August 2011 (PDT)
• Better categorization of tests, and ability to run tests by category, e.g.:
  • by CPU model
  • by ISA
  • by Ruby protocol
  • by length
  • I believe these can all be satisfied via decorators and command line options we define --Saidi 18:09, 22 August 2011 (PDT)
• 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.
  • This is future work. needed for sure, but a bit orthoginal --Saidi 18:09, 22 August 2011 (PDT)
• Better checkpoint testing
  • I believe we can create dependent tests, although it's an open question exactly how this interacts with the xdist stuff. --Saidi 18:09, 22 August 2011 (PDT)
  • some of this doesn't really depend on the regression framework, just needs new tests
  • e.g., integrating util/checkpoint-tester.py
    • I would rather duplicate/steal the functionality into a Base class for regression testing --Saidi 18:09, 22 August 2011 (PDT)
• Support for random testing (e.g., for background testing processes)
  • Random latencies?
  • Random testing a la memory testers but with different seeds, longer intervals
    • I don't think there is anything preventing us from calling random.randint() in a test, and it's all python --Saidi 18:09, 22 August 2011 (PDT)
• Decouple from SCons
  • Avoid having scons dependency bugs force unnecessary re-running of tests, particularly for update-refs
    • not quite sure how we would do updaterefs with py.test. I've been using a 4 line bash script anyway, because I don't want to update every test, only the failed ones --Saidi 18:09, 22 August 2011 (PDT)
  • 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
    • pytest should be able to do this, again, open question it if does it via multi-threads and bsub -K on each job or if we should have pytest directly connect to running machines.
• Easy support for running separate tests where only the input parameters differ
  • As long as we do a good job creating the base class, this should be trivial --Saidi 18:09, 22 August 2011 (PDT)
  • 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.
    • Big question is how do we store all of the file data, especially when the stats are python objects (hint hint nate). Perhaps putting everything in a SQLite DB might not be a bad approach so we don't need complex directory structures --Saidi 18:09, 22 August 2011 (PDT)
  • Similarly, we could/should test topologies this way as well.
• Automated way to use nightly regressions as a basis for updating "m5-stable"
  • I would vote to defer this until we have a continuous integration solution --Saidi 18:09, 22 August 2011 (PDT)
  • 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.
    • I think all of these can be wrapped up into what pass/fail means and what is printed. Depending on the test we could say that stats don't matter but simout must be the same, etc. --Saidi 18:09,

22 August 2011 (PDT)

• • It would be nice to also track macroop/regular instruction and microop counts separately. In SE mode at least, if the number of ISA level instructions changes it's more important than if the number of microops change. That way when you change microcode, you can tell if the same instructions are executing more or less and just their implementation has changed. --Gblack 21:35, 22 August 2011 (PDT)
  • For directed tests like the ones that test specific instructions, it might be worthwhile to add pseudo ops that signal success or completion through some sort of backdoor. There could be similar instructions which pause the simulation and give the simulation script a chance to look at state (I don't think we can now, but it would be handy) and decide if, say, the add put the right thing in the right registers, or memory was updated correctly. One major drawback to this sort of approach is that we couldn't easily run these on real machines. On the positive side, we could avoid problems where complementary bugs make a test pass when it shouldn't. For instance, maybe adds and subtracts have the same bug, so both the output and checking the output against reference values might be wrong.--Gblack 21:35, 22 August 2011 (PDT)

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)