def test_get_nvidia_label(self):
node_id = 6
job_id = 10
gpu_count = 2
required_gpus = 2
GPUManager.define_node_gpus(node_id, gpu_count)
GPUManager.reserve_gpus_for_job(node_id, required_gpus)
GPUManager.assign_gpus_for_job(node_id, job_id, required_gpus)
nvidia_label = GPUManager.get_nvidia_docker_label(node_id, job_id)
self.assertEqual(nvidia_label, "0,1")
gpu_count = 4
job_id = 11
GPUManager.define_node_gpus(node_id, gpu_count)
GPUManager.reserve_gpus_for_job(node_id, required_gpus)
GPUManager.assign_gpus_for_job(node_id, job_id, required_gpus)
nvidia_label = GPUManager.get_nvidia_docker_label(node_id, job_id)
self.assertEqual(nvidia_label, "2,3")
def test_calls_where_node_has_no_gpus(self):
node_id = 7
job_id = 10
gpu_count = 2
required_gpus = 2
GPUManager.define_node_gpus(node_id, gpu_count)
node_id = 8
self.assertFalse(
GPUManager.reserve_gpus_for_job(node_id, required_gpus))
self.assertFalse(
GPUManager.assign_gpus_for_job(node_id, job_id, required_gpus))
nvidia_label = GPUManager.get_nvidia_docker_label(node_id, job_id)
self.assertEqual(nvidia_label, "")
def test_release_gpu(self):
node_id = 7
job_id = 10
gpu_count = 2
required_gpus = 2
GPUManager.define_node_gpus(node_id, gpu_count)
GPUManager.reserve_gpus_for_job(node_id, required_gpus)
self.assertTrue(
GPUManager.assign_gpus_for_job(node_id, job_id, required_gpus))
job_id = 11
self.assertFalse(
GPUManager.reserve_gpus_for_job(
node_id, required_gpus)) # shouldnt have enough GPUs
GPUManager.release_gpus(node_id, 10)
self.assertTrue(GPUManager.reserve_gpus_for_job(
node_id, required_gpus)) #gpus should be avail again
self.assertTrue(
GPUManager.assign_gpus_for_job(
node_id, job_id, required_gpus)) #gpus should be avail again
nvidia_label = GPUManager.get_nvidia_docker_label(node_id, job_id)
self.assertEqual(nvidia_label, "0,1")
def _configure_all_tasks(self, config, job_exe, job_type):
"""Configures the given execution with items that apply to all tasks
:param config: The execution configuration
:type config: :class:`job.execution.configuration.json.exe_config.ExecutionConfiguration`
:param job_exe: The job execution model being scheduled
:type job_exe: :class:`job.models.JobExecution`
:param job_type: The job type model
:type job_type: :class:`job.models.JobType`
"""
config.set_task_ids(job_exe.get_cluster_id())
for task_type in config.get_task_types():
# Configure env vars describing allocated task resources
env_vars = {}
nvidia_docker_label = None
for resource in config.get_resources(task_type).resources:
env_name = 'ALLOCATED_%s' % normalize_env_var_name(
resource.name)
env_vars[
env_name] = '%.1f' % resource.value # Assumes scalar resources
if resource.name == "gpus" and int(resource.value) > 0:
gpu_list = GPUManager.get_nvidia_docker_label(
job_exe.node_id, job_exe.job_id)
nvidia_docker_label = DockerParameter(
'env', 'NVIDIA_VISIBLE_DEVICES={}'.format(
gpu_list.strip(',')))
# Configure env vars for Scale meta-data
env_vars['SCALE_JOB_ID'] = unicode(job_exe.job_id)
env_vars['SCALE_EXE_NUM'] = unicode(job_exe.exe_num)
if job_exe.recipe_id:
env_vars['SCALE_RECIPE_ID'] = unicode(job_exe.recipe_id)
if job_exe.batch_id:
env_vars['SCALE_BATCH_ID'] = unicode(job_exe.batch_id)
# Configure workspace volumes
workspace_volumes = {}
for task_workspace in config.get_workspaces(task_type):
logger.debug(self._workspaces)
workspace_model = self._workspaces[task_workspace.name]
# TODO: Should refactor workspace broker to return a Volume object and remove BrokerVolume
if workspace_model.volume:
vol_name = get_workspace_volume_name(
job_exe, task_workspace.name)
cont_path = get_workspace_volume_path(workspace_model.name)
if workspace_model.volume.host:
host_path = workspace_model.volume.remote_path
volume = Volume(vol_name,
cont_path,
task_workspace.mode,
is_host=True,
host_path=host_path)
else:
driver = workspace_model.volume.driver
driver_opts = {}
# TODO: Hack alert for nfs broker, as stated above, we should return Volume from broker
if driver == 'nfs':
driver_opts = {
'share': workspace_model.volume.remote_path
}
volume = Volume(vol_name,
cont_path,
task_workspace.mode,
is_host=False,
driver=driver,
driver_opts=driver_opts)
workspace_volumes[task_workspace.name] = volume
config.add_to_task(task_type,
env_vars=env_vars,
wksp_volumes=workspace_volumes)
# Labels for metric grouping
job_id_label = DockerParameter(
'label', 'scale-job-id={}'.format(job_exe.job_id))
job_execution_id_label = DockerParameter(
'label', 'scale-job-execution-id={}'.format(job_exe.exe_num))
job_type_name_label = DockerParameter(
'label', 'scale-job-type-name={}'.format(job_type.name))
job_type_version_label = DockerParameter(
'label', 'scale-job-type-version={}'.format(job_type.version))
main_label = DockerParameter('label', 'scale-task-type=main')
if nvidia_docker_label:
nvidia_runtime_param = DockerParameter('runtime', 'nvidia')
config.add_to_task('main',
docker_params=[
job_id_label, job_type_name_label,
job_type_version_label,
job_execution_id_label, main_label,
nvidia_docker_label, nvidia_runtime_param
])
else:
config.add_to_task('main',
docker_params=[
job_id_label, job_type_name_label,
job_type_version_label,
job_execution_id_label, main_label
])
if not job_type.is_system:
pre_label = DockerParameter('label', 'scale-task-type=pre')
post_label = DockerParameter('label', 'scale-task-type=post')
config.add_to_task('pre',
docker_params=[
job_id_label, job_type_name_label,
job_type_version_label,
job_execution_id_label, pre_label
])
config.add_to_task('post',
docker_params=[
job_id_label, job_type_name_label,
job_type_version_label,
job_execution_id_label, post_label
])
# Configure tasks for logging
if settings.LOGGING_ADDRESS is not None:
log_driver = DockerParameter('log-driver', 'fluentd')
fluent_precision = DockerParameter(
'log-opt', 'fluentd-sub-second-precision=true')
log_address = DockerParameter(
'log-opt', 'fluentd-address=%s' % settings.LOGGING_ADDRESS)
if not job_type.is_system:
pre_task_tag = DockerParameter(
'log-opt', 'tag=%s|%s|%s|%s|%s' %
(config.get_task_id('pre'), job_type.name,
job_type.version, job_exe.job_id, job_exe.exe_num))
config.add_to_task('pre',
docker_params=[
log_driver, fluent_precision,
log_address, pre_task_tag
])
post_task_tag = DockerParameter(
'log-opt', 'tag=%s|%s|%s|%s|%s' %
(config.get_task_id('post'), job_type.name,
job_type.version, job_exe.job_id, job_exe.exe_num))
config.add_to_task('post',
docker_params=[
log_driver, fluent_precision,
log_address, post_task_tag
])
# TODO: remove es_urls parameter when Scale no longer supports old style job types
# Post task needs ElasticSearch URL to grab logs for old artifact registration
es_param = DockerParameter(
'env', 'ELASTICSEARCH_URL=%s' % settings.ELASTICSEARCH_URL)
config.add_to_task('post', docker_params=[es_param])
main_task_tag = DockerParameter(
'log-opt', 'tag=%s|%s|%s|%s|%s' %
(config.get_task_id('main'), job_type.name, job_type.version,
job_exe.job_id, job_exe.exe_num))
config.add_to_task('main',
docker_params=[
log_driver, fluent_precision, log_address,
main_task_tag
])