def _test_job_arrays_and_parallel_jobs(remote_command_executor, region,
stack_name, scaledown_idletime,
max_slots):
logging.info(
"Testing cluster scales correctly with array jobs and parallel jobs")
torque_commands = TorqueCommands(remote_command_executor)
result = remote_command_executor.run_remote_command(
"echo 'sleep 30' | qsub -t 1-{0}".format(max_slots),
raise_on_error=False)
array_job_id = torque_commands.assert_job_submitted(result.stdout)
result = remote_command_executor.run_remote_command(
"echo 'sleep 30' | qsub -l nodes=2:ppn=1", raise_on_error=False)
parallel_job_id = torque_commands.assert_job_submitted(result.stdout)
# Assert scaling worked as expected
assert_scaling_worked(torque_commands,
region,
stack_name,
scaledown_idletime,
expected_max=3,
expected_final=0)
# Assert jobs were completed
for i in range(1, max_slots + 1):
_assert_job_completed(remote_command_executor,
array_job_id.replace("[]", "[{0}]".format(i)))
_assert_job_completed(remote_command_executor, parallel_job_id)
def _test_job_dependencies(remote_command_executor, region, stack_name,
scaledown_idletime):
logging.info(
"Testing cluster doesn't scale when job dependencies are not satisfied"
)
torque_commands = TorqueCommands(remote_command_executor)
result = torque_commands.submit_command("sleep 60", nodes=1)
job_id = torque_commands.assert_job_submitted(result.stdout)
result = remote_command_executor.run_remote_command(
"echo 'sleep 1' | qsub -W depend=afterok:{0}".format(job_id),
raise_on_error=False)
dependent_job_id = torque_commands.assert_job_submitted(result.stdout)
assert_that(_get_job_state(remote_command_executor,
dependent_job_id)).is_equal_to("H")
# Assert scaling worked as expected
assert_scaling_worked(torque_commands,
region,
stack_name,
scaledown_idletime,
expected_max=1,
expected_final=0)
# Assert jobs were completed
_assert_job_completed(remote_command_executor, job_id)
_assert_job_completed(remote_command_executor, dependent_job_id)
def _test_non_runnable_jobs(remote_command_executor, max_queue_size, max_slots,
region, cluster, scaledown_idletime):
logging.info("Testing jobs that violate scheduling requirements")
torque_commands = TorqueCommands(remote_command_executor)
# Make sure the cluster has at least 1 node in the queue so that we can verify cluster scales down correctly
if torque_commands.compute_nodes_count() == 0:
result = torque_commands.submit_command("sleep 1")
job_id = torque_commands.assert_job_submitted(result.stdout)
torque_commands.wait_job_completed(job_id)
assert_that(torque_commands.compute_nodes_count()).is_greater_than(0)
logging.info(
"Testing cluster doesn't scale when job requires a capacity that is higher than the max available"
)
# nodes limit enforced by scheduler
result = remote_command_executor.run_remote_command(
"echo 'sleep 1000' | qsub -l nodes={0}".format(max_queue_size + 1),
raise_on_error=False)
assert_that(result.stdout).contains("Job exceeds queue resource limits")
# ppn limit enforced by daemons
result = remote_command_executor.run_remote_command(
"echo 'sleep 1000' | qsub -l nodes=1:ppn={0}".format(max_slots + 1),
raise_on_error=False)
ppn_job_id = torque_commands.assert_job_submitted(result.stdout)
# ppn total limit enforced by scheduler
result = remote_command_executor.run_remote_command(
"echo 'sleep 1000' | qsub -l nodes=1:ppn={0}".format((max_slots *
max_queue_size) +
1),
raise_on_error=False)
assert_that(result.stdout).contains("Job exceeds queue resource limits")
# ncpus limit enforced by scheduler
result = remote_command_executor.run_remote_command(
"echo 'sleep 1000' | qsub -l ncpus={0}".format(max_slots + 1),
raise_on_error=False)
assert_that(result.stdout).contains("Job exceeds queue resource limits")
logging.info("Testing cluster doesn't scale when job is set on hold")
result = remote_command_executor.run_remote_command(
"echo 'sleep 1000' | qsub -l nodes=1 -h", raise_on_error=False)
hold_job_id = torque_commands.assert_job_submitted(result.stdout)
logging.info(
"Testing cluster scales down when pending jobs cannot be submitted")
assert_scaling_worked(torque_commands,
region,
cluster.cfn_name,
scaledown_idletime,
expected_max=1,
expected_final=0)
# Assert jobs are still pending
assert_that(_get_job_state(remote_command_executor,
ppn_job_id)).is_equal_to("Q")
assert_that(_get_job_state(remote_command_executor,
hold_job_id)).is_equal_to("H")
def _test_dynamic_cluster_limits(remote_command_executor, max_queue_size, max_slots, region, asg_name):
logging.info("Testing cluster limits are dynamically updated")
torque_commands = TorqueCommands(remote_command_executor)
# Make sure cluster is scaled to 0 when this test starts
assert_that(torque_commands.compute_nodes_count()).is_equal_to(0)
_assert_scheduler_configuration(remote_command_executor, torque_commands, max_slots, max_queue_size)
# Submit a job to scale up to 1 node
result = torque_commands.submit_command("sleep 1", nodes=1)
job_id = torque_commands.assert_job_submitted(result.stdout)
# Change ASG max size
asg_client = boto3.client("autoscaling", region_name=region)
new_max_size = max_queue_size + 1
asg_client.update_auto_scaling_group(AutoScalingGroupName=asg_name, MaxSize=new_max_size)
# sleeping for 200 seconds since daemons fetch this data every 3 minutes
time.sleep(200)
# Wait for job completion to be sure cluster scaled
torque_commands.wait_job_completed(job_id)
_assert_scheduler_configuration(remote_command_executor, torque_commands, max_slots, new_max_size)
# Restore initial cluster size
asg_client.update_auto_scaling_group(AutoScalingGroupName=asg_name, MaxSize=max_queue_size)
# sleeping for 200 seconds since daemons fetch this data every 3 minutes
time.sleep(200)
_assert_scheduler_configuration(remote_command_executor, torque_commands, max_slots, max_queue_size)
def _test_jobs_executed_concurrently(remote_command_executor, max_slots):
logging.info("Testing jobs are executed concurrently and nodes are fully allocated")
torque_commands = TorqueCommands(remote_command_executor)
# GIVEN: a cluster with 3 free nodes
assert_that(torque_commands.compute_nodes_count()).is_equal_to(3)
# WHEN: an array job that requires 3 nodes and all slots is submitted
jobs_start_time = int(remote_command_executor.run_remote_command("date +%s").stdout)
job_exec_time = 30
job_ids = []
for i in range(0, 3 * max_slots):
result = torque_commands.submit_command(
f"sleep {job_exec_time} && hostname > /shared/job{i} && date +%s >> /shared/end_time", nodes=1, slots=1
)
job_id = torque_commands.assert_job_submitted(result.stdout)
job_ids.append(job_id)
# THEN: cluster scales down correctly after completion
watch_compute_nodes(torque_commands, minutes(10), 0)
for id in job_ids:
_assert_job_completed(remote_command_executor, id)
# THEN: each host executes 4 jobs in the expected time
jobs_to_hosts_count = (
remote_command_executor.run_remote_command("cat /shared/job* | sort | uniq -c | awk '{print $1}'")
.stdout.strip()
.splitlines()
)
assert_that(jobs_to_hosts_count).is_equal_to(["4", "4", "4"])
# verify execution time
jobs_completion_time = int(
remote_command_executor.run_remote_command("cat /shared/end_time | sort -n | tail -1").stdout.split()[-1]
)
assert_that(jobs_completion_time - jobs_start_time).is_greater_than(0).is_less_than(2 * job_exec_time)
def _test_torque_job_submit(remote_command_executor, test_datadir):
"""Test torque job submit command in slurm cluster."""
logging.info("Testing cluster submits job by torque command")
torque_commands = TorqueCommands(remote_command_executor)
result = torque_commands.submit_script(str(test_datadir / "torque_job.sh"))
torque_commands.assert_job_submitted(result.stdout)