def launcher(request):
if request.param == 'launcherwrapper':
# We set the registered_name for the LauncherWrapper just for
# convenience for the rest of the unit tests
wrapper_cls = launchers.LauncherWrapper
wrapper_cls.registered_name = 'launcherwrapper'
return wrapper_cls(getlauncher('alps')(), 'ddt', ['--offline'])
return getlauncher(request.param)()
def prepare_run(self):
if self.current_partition.fullname in ['daint:gpu', 'dom:gpu']:
num_workers = 12
exec_cores = 3
else:
num_workers = 36
exec_cores = 9
self.variables = {
'SPARK_WORKER_CORES': str(num_workers),
'SPARK_LOCAL_DIRS': '"/tmp"',
}
self.executable = 'spark-submit'
self.executable_opts = [
f'--conf spark.default.parallelism={num_workers}',
f'--conf spark.executor.cores={exec_cores}',
f'--conf spark.executor.memory=15g', f'--master $SPARKURL'
]
if self.variant == 'spark':
self.executable_opts += [
'--class org.apache.spark.examples.SparkPi',
'$EBROOTSPARK/examples/jars/spark-examples*.jar 10000'
]
else:
self.executable_opts.append('spark_pi.py')
# The job launcher has to be changed since the `spark-submit`
# script is not used with srun.
self.job.launcher = getlauncher('local')()
def _setup_fake_check():
# A bit hacky, but we don't want to run a full test every time
test = _FakeCheck()
test._job = Job.create(
getscheduler('local')(),
getlauncher('local')(), 'fakejob')
test.job._completion_time = time.time()
return test
def fake_check():
class _FakeCheck(rfm.RegressionTest):
pass
# A bit hacky, but we don't want to run a full test every time
test = _FakeCheck()
test._job = Job.create(
getscheduler('local')(),
getlauncher('local')(), 'fakejob')
test.job._completion_time = time.time()
return test
def _make_flexible_job(flex_type, **jobargs):
ret = Job.create(slurm_scheduler_patched,
getlauncher('local')(),
name='testjob',
workdir=tmp_path,
script_filename=str(tmp_path / 'job.sh'),
stdout=str(tmp_path / 'job.out'),
stderr=str(tmp_path / 'job.err'),
sched_flex_alloc_nodes=flex_type,
**jobargs)
ret.num_tasks = 0
ret.num_tasks_per_node = 4
return ret
def set_profiler(self):
# mpirun ko: orte_ess_init failed
# --> Returned value No permission (-17) instead of ORTE_SUCCESS
# [mpirun -np 1 <self.job.launcher.options>] ./myexe ...
# https://github.com/open-mpi/ompi/issues/6981
# https://github.com/open-mpi/ompi/issues/8017
self.job.launcher = getlauncher('local')()
self.executable = 'rocprof'
self.executable_opts += [
'-i', self.metric_file, '--timestamp', 'on', '--basenames', 'on',
f'{self.exe_path}/sedov-cuda', f"-n {self.cubeside}",
f"-s {self.steps}"
]
def setup_per_env(self):
envname = self.current_environ.name
self.prerun_cmds = [
f'source ./shrc', f'mv {self.configs[envname]} config'
]
self.executable_opts = [
f'--config={self.configs[envname]}', '--platform NVIDIA',
'--tune=base', '--device GPU'
] + self.benchmarks[envname]
self.reference = {
'dom:gpu': self.refs[envname],
'daint:gpu': self.refs[envname]
}
self.sanity_patterns = self.sanity_patterns_[envname]
self.perf_patterns = self.perf_patterns_[envname]
# The job launcher has to be changed since the `runspec`
# script is not used with srun.
self.job.launcher = getlauncher('local')()
def prepare_run(self):
if self.current_partition.fullname in ['daint:gpu', 'dom:gpu']:
num_workers = 12
exec_cores = 3
else:
num_workers = 36
exec_cores = 9
self.variables = {
'SPARK_WORKER_CORES': '%s' % num_workers,
'SPARK_LOCAL_DIRS': '"/tmp"',
}
self.executable = (
'spark-submit --conf spark.default.parallelism=%s '
'--conf spark.executor.cores=%s --conf spark.executor.memory=15g '
'--master $SPARKURL --class org.apache.spark.examples.SparkPi '
'$EBROOTSPARK/examples/jars/spark-examples_2.11-2.3.1.jar 10000;' %
(num_workers, exec_cores))
# The job launcher has to be changed since the `spark-submit`
# script is not used with srun.
self.job.launcher = getlauncher('local')()
def fake_check():
class _FakeCheck(rfm.RegressionTest):
pass
@sn.deferrable
def error():
raise BaseException
# A bit hacky, but we don't want to run a full test every time
test = _FakeCheck()
test._job = Job.create(
getscheduler('local')(),
getlauncher('local')(), 'fakejob')
test.job._completion_time = time.time()
test.job._jobid = 12345
test.job._nodelist = ['localhost']
test.custom = 'hello extras'
test.custom_list = ['custom', 3.0, ['hello', 'world']]
test.custom_dict = {'a': 1, 'b': 2}
test.deferred = sn.defer('hello')
test.deferred_error = error()
return test
def fake_check():
class _FakeCheck(rfm.RegressionTest):
param = parameter(range(3), loggable=True, fmt=lambda x: 10 * x)
custom = variable(str, value='hello extras', loggable=True)
custom_list = variable(list,
value=['custom', 3.0, ['hello', 'world']],
loggable=True)
custom_dict = variable(dict, value={'a': 1, 'b': 2}, loggable=True)
# x is a variable that is loggable, but is left undefined. We want to
# make sure that logging does not crash and simply reports is as
# undefined
x = variable(str, loggable=True)
# A bit hacky, but we don't want to run a full test every time
test = _FakeCheck(variant_num=1)
test._job = Job.create(
getscheduler('local')(),
getlauncher('local')(), 'fakejob')
test.job._completion_time = time.time()
test.job._jobid = 12345
test.job._nodelist = ['localhost']
return test
def launcher(self):
return getlauncher('srunalloc')()
def create(cls, site_config):
# Create the whole system hierarchy from bottom up
sysname = site_config.get('systems/0/name')
partitions = []
config_save = site_config.subconfig_system
for p in site_config.get('systems/0/partitions'):
site_config.select_subconfig(f'{sysname}:{p["name"]}')
partid = f"systems/0/partitions/@{p['name']}"
part_name = site_config.get(f'{partid}/name')
part_sched = getscheduler(site_config.get(f'{partid}/scheduler'))
part_launcher = getlauncher(site_config.get(f'{partid}/launcher'))
part_container_environs = {}
for i, p in enumerate(
site_config.get(f'{partid}/container_platforms')):
ctype = p['type']
part_container_environs[ctype] = Environment(
name=f'__rfm_env_{ctype}',
modules=site_config.get(
f'{partid}/container_platforms/{i}/modules'),
variables=site_config.get(
f'{partid}/container_platforms/{i}/variables'))
part_environs = [
ProgEnvironment(
name=e,
modules=site_config.get(f'environments/@{e}/modules'),
variables=site_config.get(f'environments/@{e}/variables'),
cc=site_config.get(f'environments/@{e}/cc'),
cxx=site_config.get(f'environments/@{e}/cxx'),
ftn=site_config.get(f'environments/@{e}/ftn'),
cppflags=site_config.get(f'environments/@{e}/cppflags'),
cflags=site_config.get(f'environments/@{e}/cflags'),
cxxflags=site_config.get(f'environments/@{e}/cxxflags'),
fflags=site_config.get(f'environments/@{e}/fflags'),
ldflags=site_config.get(f'environments/@{e}/ldflags'))
for e in site_config.get(f'{partid}/environs')
]
partitions.append(
SystemPartition(
parent=site_config.get('systems/0/name'),
name=part_name,
sched_type=part_sched,
launcher_type=part_launcher,
descr=site_config.get(f'{partid}/descr'),
access=site_config.get(f'{partid}/access'),
resources=site_config.get(f'{partid}/resources'),
environs=part_environs,
container_environs=part_container_environs,
local_env=Environment(
name=f'__rfm_env_{part_name}',
modules=site_config.get(f'{partid}/modules'),
variables=site_config.get(f'{partid}/variables')),
max_jobs=site_config.get(f'{partid}/max_jobs'),
prepare_cmds=site_config.get(f'{partid}/prepare_cmds'),
processor=site_config.get(f'{partid}/processor'),
devices=site_config.get(f'{partid}/devices'),
extras=site_config.get(f'{partid}/extras')))
# Restore configuration, but ignore unresolved sections or
# configuration parameters at the system level; if we came up to this
# point, then all is good at the partition level, which is enough.
site_config.select_subconfig(config_save, ignore_resolve_errors=True)
return System(
name=sysname,
descr=site_config.get('systems/0/descr'),
hostnames=site_config.get('systems/0/hostnames'),
modules_system=site_config.get('systems/0/modules_system'),
preload_env=Environment(
name=f'__rfm_env_{sysname}',
modules=site_config.get('systems/0/modules'),
variables=site_config.get('systems/0/variables')),
prefix=site_config.get('systems/0/prefix'),
outputdir=site_config.get('systems/0/outputdir'),
resourcesdir=site_config.get('systems/0/resourcesdir'),
stagedir=site_config.get('systems/0/stagedir'),
partitions=partitions)
def set_launcher(self):
# The job launcher has to be changed since the `start_analytics`
# script is not used with srun.
self.job.launcher = getlauncher('local')()
def launcher(self):
return getlauncher('ssh')()
def launcher(self):
return launchers.LauncherWrapper(
getlauncher('alps')(), 'ddt', ['--offline']
)
def launcher(self):
return getlauncher('local')()
def set_launcher(self):
# The job launcher has to be changed because
# the tool can be called without srun
self.job.launcher = getlauncher('local')()
def create(cls, site_config):
# Create the whole system hierarchy from bottom up
sysname = site_config.get('systems/0/name')
partitions = []
config_save = site_config.subconfig_system
for p in site_config.get('systems/0/partitions'):
site_config.select_subconfig(f'{sysname}:{p["name"]}')
partid = f"systems/0/partitions/@{p['name']}"
part_name = site_config.get(f'{partid}/name')
part_sched = getscheduler(site_config.get(f'{partid}/scheduler'))
part_launcher = getlauncher(site_config.get(f'{partid}/launcher'))
part_container_environs = {}
for i, p in enumerate(
site_config.get(f'{partid}/container_platforms')):
ctype = p['type']
part_container_environs[ctype] = Environment(
name=f'__rfm_env_{ctype}',
modules=site_config.get(
f'{partid}/container_platforms/{i}/modules'),
variables=site_config.get(
f'{partid}/container_platforms/{i}/variables'))
part_environs = [
ProgEnvironment(
name=e,
modules=site_config.get(f'environments/@{e}/modules'),
variables=site_config.get(f'environments/@{e}/variables'),
cc=site_config.get(f'environments/@{e}/cc'),
cxx=site_config.get(f'environments/@{e}/cxx'),
ftn=site_config.get(f'environments/@{e}/ftn'),
cppflags=site_config.get(f'environments/@{e}/cppflags'),
cflags=site_config.get(f'environments/@{e}/cflags'),
cxxflags=site_config.get(f'environments/@{e}/cxxflags'),
fflags=site_config.get(f'environments/@{e}/fflags'),
ldflags=site_config.get(f'environments/@{e}/ldflags'))
for e in site_config.get(f'{partid}/environs')
]
partitions.append(
SystemPartition(
parent=site_config.get('systems/0/name'),
name=part_name,
scheduler=part_sched,
launcher=part_launcher,
descr=site_config.get(f'{partid}/descr'),
access=site_config.get(f'{partid}/access'),
resources=site_config.get(f'{partid}/resources'),
environs=part_environs,
container_environs=part_container_environs,
local_env=Environment(
name=f'__rfm_env_{part_name}',
modules=site_config.get(f'{partid}/modules'),
variables=site_config.get(f'{partid}/variables')),
max_jobs=site_config.get(f'{partid}/max_jobs')))
# Restore configuration
site_config.select_subconfig(config_save)
return System(
name=sysname,
descr=site_config.get('systems/0/descr'),
hostnames=site_config.get('systems/0/hostnames'),
modules_system=site_config.get('systems/0/modules_system'),
preload_env=Environment(
name=f'__rfm_env_{sysname}',
modules=site_config.get('systems/0/modules'),
variables=site_config.get('systems/0/variables')),
prefix=site_config.get('systems/0/prefix'),
outputdir=site_config.get('systems/0/outputdir'),
resourcesdir=site_config.get('systems/0/resourcesdir'),
stagedir=site_config.get('systems/0/stagedir'),
partitions=partitions)
def set_job_launcher(self):
# The job launcher has to be changed since the `spark-submit`
# script is not used with srun.
self.job.launcher = getlauncher('local')()
def prepare_run(self):
# Change the job launcher since `ipython`
# needs to be launched without `srun`.
self.job.launcher = getlauncher('local')()
def set_launcher(self):
# The job launcher has to be changed because
# valgrind4hpc can not be called with srun
self.job.launcher = getlauncher('local')()
def launcher():
return getlauncher('local')
def launcher(self):
return getlauncher('alps')()
def set_launcher(self):
# The job launcher has to be changed to local since greasy
# make calls to srun
self.job.launcher = getlauncher('local')()
def launcher(self):
return getlauncher('mpirun')()
def launcher(self):
return getlauncher('mpiexec')()
def launcher(self):
return getlauncher('upcxx-run')()
