[WIP] OSU Microbenchmark example for LUMI #1
Conversation
olouant
changed the title
| stack_name = os.getenv('LUMI_STACK_NAME', None) | ||
| stack_version = os.getenv('LUMI_STACK_VERSION', None) | ||
|
|
||
| environs = ['PrgEnv-gnu', 'PrgEnv-cray'] | ||
|
|
||
| if stack_name and stack_name == 'LUMI': | ||
| environs += ['cpeGNU', 'cpeCray'] | ||
|
|
||
| if version.parse(stack_version) > version.parse('21.08'): | ||
| environs += ['cpeAOCC'] |
There was a problem hiding this comment.
I add to create this hacky solution in order to support the different software stack of LUMI. cpeAOCC has been introduced by LUMI/21.12 and is not present in 21.08. The PrgEnv-aocc provided by Cray is still broken.
config/lumi.py
Outdated
| ], | ||
| 'max_jobs': 100, | ||
| 'modules': ['LUMI'], |
There was a problem hiding this comment.
If we do this, the default module (21.08) is loaded and we do not test 21.12. It also lead to failure of the cpeAOCC tests.
There was a problem hiding this comment.
I think that removing 'modules': ['LUMI'] is not a good idea. To load the cpeXXX modules, the LUMI module needs to be loaded. ReFrame does that so we don't have to do it ourselves. A solution can be to keep LUMI with no version on the configuration and then use the --map-module LUMI:LUMI/21.12 when we run the tests.
There was a problem hiding this comment.
make sense. I restored 'modules': ['LUMI'] in the configuration
There was a problem hiding this comment.
It may be me removing LUMI by accident, apologize.
| class osu_init(osu_benchmark_test_base): | ||
| @sanity_function | ||
| def validate_test(self): | ||
| #return sn.assert_eq(self.job.exitcode, 0) | ||
| return sn.assert_found(rf'^nprocs: {self.num_tasks}', self.stdout) |
There was a problem hiding this comment.
I have added osu_init test, just as an exercise. Please review if syntax is correct and I am not disruptive to generic class.
|
Yesterday Kurt made me realise tests should be build using partition |
Not sure if it is the best approach for this - PR #3 |
| self.build_system.cxx = 'hipcc' | ||
| self.build_system.cflags = ['-I/opt/cray/pe/mpich/8.1.8/ofi/crayclang/10.0/include'] | ||
| self.build_system.ldflags = ['-L/opt/cray/pe/mpich/8.1.8/ofi/crayclang/10.0/lib', '-lmpi', '-L/opt/cray/pe/mpich/8.1.8/gtl/lib', '-lmpi_gtl_hsa'] | ||
| self.build_system.cppflags = ['-D__HIP_PLATFORM_AMD__'] |
There was a problem hiding this comment.
Would it make sense something like what suggested here #4 (comment)?
|
Waiting for #3 to be merged, cleanup and catch up with reframe upstream required |
This is an example with some OSU tests based on a test from
reframe/hpctestlib. This branch is the continuation of the one from the PR reframe-hpc#2421 which still hasn't been finished so this is going to change.The tests can be run with
from the reframe base directory.