[test] Make CUDA kernel latency tests generic

[vkarak]

The idea of making a test generic is to make no assumption as of what is self.current_system. This essentially means the following:

1. If you check against self.current_system or self.current_partition or self.current_environ, make sure you handle the case of an unknown system or environment.
2. If you are running on an unknown system, make sure to include * in self.valid_systems and self.valid_prog_environs.
3. If you are on a performance test, make sure that a * entry exists inside the references dictionary. Its value should be (0, None, None, 'unit').

Addresses UES-339

[jgphpc]

Modyfing config
to have cuda/10.1 instead of cuda will fix UES-339

It would help me test this PR.

[jgphpc]

I get on ault:

Found 4 gpu(s)
[gpu 0] Kernel launch latency: 8.67954 us
[gpu 1] Kernel launch latency: 9.11042 us
[gpu 2] Kernel launch latency: 8.99908 us
[gpu 3] Kernel launch latency: 8.63197 us

  * Failing phase: sanity
  * Reason: sanity error: 8 != 0

where 'latency': (0.0, None, None, 'us')

Should sanity pass or not ?

[vkarak]

@jgphpc There two things here:

1. There is a bug in my implementation, because I should have set self.num_gpus_per_node = 1.
2. This wouldn't prevent the test from failing though on an unknown system, but this is not a problem. You can't have a really generic test without any sort of hardware detection. The important things is that the test is loaded correctly and runs through on an unknown system. Then you inspect possible failures.

[teojgo]

@jgphpc Also it is a flexible test therefore it will run on all idle nodes unless instructed with --flex-alloc-tasks

[jgphpc]

Adding self.num_gpus_per_node = 4 makes the sanity test to pass, thanks @teojgo

./reframe.py --flex-alloc-tasks=1 -C config/cscs.py --system ault:intelv100 ...

[vkarak]

@jgphpc Yes, but you don't know that if running on an unknown system. Most probably the test will fail. At least it's already ready to run. That's the point.

[vkarak]

Modyfing config
to have cuda/10.1 instead of cuda will fix UES-339

It would help me test this PR.

@jgphpc Can you do a separate PR for this? Or open an issue?

[pep8speaks]

Hello @vkarak, Thank you for updating!

• In the file cscs-checks/microbenchmarks/kernel_latency/kernel_latency.py:

Line 74:1: W293 blank line contains whitespace
Line 88:1: W293 blank line contains whitespace

Do see the ReFrame Coding Style Guide

[vkarak]

@jenkins-cscs retry daint kesch

[vkarak]

@jenkins-cscs retry none