More detailed information about dependencies

[victorusu]

In one of my tests, reframe reports the -L reports

- BTest:
    Dependencies:
      ATest
...
    Valid environments:
      builtin, PrgEnv-gnu

    Valid systems:
      daint:gpu, dom:gpu, daint:mc, dom:mc

But my BTest only depends on ATest when

• The ATest has partition login and programming environment PrgEnv-gnu; and
• The BTest has partition gpu and programming environment builtin or PrgEnv-gnu

I have the following seam in my BTest class

        self.valid_systems = ['daint:gpu','dom:gpu', 'daint:mc','dom:mc']
        self.valid_prog_environs = ['builtin', 'PrgEnv-gnu', 'PrgEnv-intel']
...
        subdeps = {('gpu', 'builtin'): [('login', environment)]}
        subdeps[('gpu', 'PrgEnv-gnu')] = [('login', environment)]
        self.depends_on('ATest', how=rfm.DEPEND_EXACT, subdeps=subdeps)
...

Note: My test is based on PR #1481.

I wanted to know in which conditions BTest depends on the ATest test. I would love to see something like the following (or better)

- BTest:
    Dependencies:
      ATest('login', 'PrgEnv-gnu') when ('gpu', 'builtin')
      ATest('login', 'PrgEnv-gnu') when ('gpu', 'PrgEnv-gnu')
...
    Valid environments:
      builtin, PrgEnv-gnu, PrgEnv-intel

    Valid systems:
      daint:gpu, dom:gpu, daint:mc, dom:mc

[vkarak]

That's a valid request. Perhaps it's not gonna be reported as you propose exactly, but we should provide more detailed information about the exact dependencies.

[ekouts]

I wanted to know in which conditions BTest depends on the ATest test. I would love to see something like the following (or better)

- BTest:
    Dependencies:
      ATest('login', 'PrgEnv-gnu') when ('gpu', 'builtin')
      ATest('login', 'PrgEnv-gnu') when ('gpu', 'PrgEnv-gnu')
...
    Valid environments:
      builtin, PrgEnv-gnu, PrgEnv-intel

    Valid systems:
      daint:gpu, dom:gpu, daint:mc, dom:mc

@vkarak This is doable, but the code as it is now it prints information about the checks (RegressionTest objects). For example valid systems can be just *, instead of a list of the ones that apply on this system. Should we make the listing more detailed in general and get the information from the testcases (TestCase objects)?

[vkarak]

Couldn't you print the whole list of micro-dependencies? We have the test cases at this point. Something like that:

- B:
    Dependencies:
      ('B', 'gpu', 'builtin') -> ('A', 'login', 'gnu')
      ('B', 'gpu', 'gnu') -> ('A', 'login', 'gnu')

I'm fine to have a different notation for the test DAG vertices, e.g., B['gpu', 'builtin'] but is should be consistent everywhere. Also, I think we should always prepend the current system in the partition, because this will make it easier for the user understand, e.g., ('B', 'daint:gpu', 'builtin').

[ekouts]

Couldn't you print the whole list of micro-dependencies? We have the test cases at this point.

Yes, we have the testcases at that point. I was asking if you think it is better to get all the information from test cases, and not print for example:

Valid systems:
      *

but print the actual partitions that it will run on:

Valid systems:
      dom:mc, dom:gpu

I'm fine to have a different notation for the test DAG vertices, e.g., B['gpu', 'builtin'] but is should be consistent everywhere. Also, I think we should always prepend the current system in the partition, because this will make it easier for the user understand, e.g., ('B', 'daint:gpu', 'builtin').

I will go with ('B', 'dom:gpu', 'builtin') -> ('A', 'dom:login', 'gnu') since this is what we are using everywhere else.