[test] Enable IOR test on Daint

[GiuseppeLoRe]

No description provided.

[pep8speaks]

Hello @GiuseppeLoRe, Thank you for updating!

Cheers! There are no PEP8 issues in this Pull Request!Do see the ReFrame Coding Style Guide

Comment last updated at 2019-09-16 14:49:27 UTC

[jenkins-cscs]

Can I test this patch?

[teojgo]

ok to test

[teojgo]

You can remove the super().__init__() call in line 10. It is not needed anymore. Furthermore, remove snx1600 from the parameterized tests since you removed it from the filesystems you check.

[codecov-io]

Codecov Report

Merging #937 into master will not change coverage.
The diff coverage is n/a.

@@          Coverage Diff           @@
##           master    #937   +/-   ##
======================================
  Coverage    91.7%   91.7%           
======================================
  Files          79      79           
  Lines       10486   10486           
======================================
  Hits         9616    9616           
  Misses        870     870

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7024c0a...262e4c1. Read the comment docs.

[vkarak]

lgtm

[vkarak]

Have you tried that on Daint? For me, some tests fail. Also I think that the numbers reported by one of the read tests are skewed due to caching.

[GiuseppeLoRe]

Without the line
super().init() in IorCheck.init
I get this error

AttributeError: IorWriteCheck object has no attribute '_current_partition'

[teojgo]

@GiuseppeLoRe You can remove the super().__init()__ only in the base class that inherits the rfm.RegressionTest.

[vkarak]

@teojgo Perhaps @GiuseppeLoRe uses an older version of reframe. This feature was introduced in 2.18 2.19.

[GiuseppeLoRe]

I am using the daint default one, 2.17.1

[GiuseppeLoRe]

Should I test with 2.18 or 2.19?

[vkarak]

So then keep the call to __init__(), because we are going to use this test on Daint immediately.

[GiuseppeLoRe]

ok

[GiuseppeLoRe]

Now that I do the tests twice, one for the gpu partition and then again for the mc one, I get the caching effect on the second test.
Is there a variable "current_partition" I can use to differentiate the output files?

[vkarak]

@GiuseppeLoRe Yes, you can. This information is available in the setup() method, so you will have to override this in your test. Check the example here on how to do this. The link refers to reframe 2.17, which is the version you are using.

[GiuseppeLoRe]

@vkarak thanks for the link, but at least for me the example is not very clear.

[vkarak]

@GiuseppeLoRe What do you want to do based on the current programming environment?

[GiuseppeLoRe]

@vkarak I would like the partition mane (mc or gpu) to build the ior output file name. And thought that I could use something like https://reframe-hpc.readthedocs.io/en/stable/reference.html#reframe.core.pipeline.RegressionTest.current_partition....

[vkarak]

Yes, you can do that, but current_partition is only defined after the setup stage of the pipeline. You can do it as follows:

    def setup(self, partition, environ, **job_opts):
        super().setup(partition, environ, **job_opts)
        if self.current_partition.name == 'gpu':
            self.executable_opts = [...]
        else:
            self.executable_opts = [...]

You may also write this as follows:

    def setup(self, partition, environ, **job_opts):
        if partition.name == 'gpu':
            self.executable_opts = [...]
        else:
            self.executable_opts = [...]

        super().setup(partition, environ, **job_opts)

And from version 2.20 (or even from the current master), you will be able to write this as follows:

    @rfm.run_after('setup')
    def choose_your_function_name(self):
        if self.current_partition.name == 'gpu':
            self.executable_opts = [...]
        else:
            self.executable_opts = [...]    

[GiuseppeLoRe]

Tested on Daint.

[vkarak]

Overall lgtm now. My only question is whether it would make sense to run the test for /users from the login nodes rather than the compute nodes, so as to avoid the wait times.

[vkarak]

@GiuseppeLoRe The test is MPI-based, so ignore my comment. I am merging this.