openstack/ironic-python-agent
Code Issues Proposed changes

Merge "Add support for burnin-gpu"

Browse Source
This commit is contained in:
Zuul 2025-01-29 19:20:10 +00:00 committed by Gerrit Code Review
commit 0c35e7e2da
5 changed files with 199 additions and 44 deletions

165
ironic_python_agent/burnin.py

@ -28,71 +28,79 @@ NETWORK_BURNIN_ROLES = frozenset(['writer', 'reader'])
NETWORK_READER_CYCLE = 30
def stress_ng_cpu(node):
"""Burn-in the CPU with stress-ng
def stress_ng(node, stressor_type, default_timeout=86400):
"""Run stress-ng for different stressor types
Run stress-ng on a configurable number of CPUs for
a configurable amount of time. Without config use
all CPUs and stress them for 24 hours.
Burn-in a configurable number of CPU/VM with stress-ng,
for a configurable amount of time but default of 24 hours.
:param node: Ironic node object
:param stressor_type: 'cpu' or 'vm'
:param default_timeout: Default timeout in seconds (default: 86400)
:raises: ValueError if an unknown stressor_type is provided
:raises: CommandExecutionError if the execution of stress-ng fails.
"""
stressor_type = stressor_type.lower()
if stressor_type not in ['cpu', 'vm']:
raise ValueError("Unknown stressor type: %s" % stressor_type)
info = node.get('driver_info', {})
cpu = info.get('agent_burnin_cpu_cpu', 0)
timeout = info.get('agent_burnin_cpu_timeout', 86400)
outputfile = info.get('agent_burnin_cpu_outputfile', None)
stressor_suffix = {'cpu': 'cpu'}
if stressor_type == 'vm':
count_key = 'agent_burnin_%s_vm' % stressor_type
bytes_key = 'agent_burnin_%s_vm-bytes' % stressor_type
else:
count_key = 'agent_burnin_%s_%s' % (stressor_type,
stressor_suffix[stressor_type])
bytes_key = None
timeout_key = 'agent_burnin_%s_timeout' % stressor_type
outputfile_key = 'agent_burnin_%s_outputfile' % stressor_type
count = info.get(count_key, 0)
timeout = info.get(timeout_key, default_timeout)
outputfile = info.get(outputfile_key)
args = ['stress-ng', '--%s' % stressor_type, count, '--timeout', timeout]
if stressor_type == 'vm':
vm_bytes = info.get(bytes_key, '98%')
args.extend(['--vm-bytes', vm_bytes])
args.extend(['--metrics-brief'])
args = ('stress-ng', '--cpu', cpu, '--timeout', timeout,
'--metrics-brief')
if outputfile:
args += ('--log-file', outputfile,)
args.extend(['--log-file', outputfile])
LOG.debug('Burn-in stress_ng_cpu command: %s', args)
LOG.debug('Burn-in stress_ng_%s command: %s', stressor_type, args)
try:
_, err = utils.execute(*args)
# stress-ng reports on stderr only
LOG.info(err)
except (processutils.ProcessExecutionError, OSError) as e:
error_msg = "stress-ng (cpu) failed with error %s" % e
error_msg = 'stress-ng (%s) failed with error %s' % (stressor_type, e)
LOG.error(error_msg)
raise errors.CommandExecutionError(error_msg)
def stress_ng_cpu(node):
"""Burn-in the CPU with stress-ng"""
stress_ng(node, 'cpu')
def stress_ng_vm(node):
"""Burn-in the memory with the vm stressor in stress-ng
"""Burn-in the memory with the vm stressor in stress-ng.
Run stress-ng with a configurable number of workers on
a configurable amount of the available memory for
a configurable amount of time. Without config use
as many workers as CPUs, 98% of the memory and stress
it for 24 hours.
:param node: Ironic node object
:raises: CommandExecutionError if the execution of stress-ng fails.
"""
info = node.get('driver_info', {})
vm = info.get('agent_burnin_vm_vm', 0)
vm_bytes = info.get('agent_burnin_vm_vm-bytes', '98%')
timeout = info.get('agent_burnin_vm_timeout', 86400)
outputfile = info.get('agent_burnin_vm_outputfile', None)
args = ('stress-ng', '--vm', vm, '--vm-bytes', vm_bytes,
'--timeout', timeout, '--metrics-brief')
if outputfile:
args += ('--log-file', outputfile,)
LOG.debug('Burn-in stress_ng_vm command: %s', args)
try:
_, err = utils.execute(*args)
# stress-ng reports on stderr only
LOG.info(err)
except (processutils.ProcessExecutionError, OSError) as e:
error_msg = "stress-ng (vm) failed with error %s" % e
LOG.error(error_msg)
raise errors.CommandExecutionError(error_msg)
stress_ng(node, 'vm')
def _smart_test_status(device):
@ -420,3 +428,84 @@ def fio_network(node):
irole = "reader" if (role == "writer") else "writer"
logfilename = outputfile + '.' + irole
_do_fio_network(not role == 'writer', runtime, partner, logfilename)
def _gpu_burn_check_count(install_dir, count):
"""Check the count of GPUs with gpu-burn
Run a check to confirm how many GPUs are seen by the OS.
:param install_dir: The location where gpu-burn has been installed.
:param count: The number of expected GPUs.
:raises: CleaningError if the incorrect number of GPUs found.
:raises: CommandExecutionError if the execution of gpu-burn fails.
"""
args = ['./gpu_burn', '-l']
LOG.debug('Burn-in gpu count command: %s', args)
try:
out, _ = utils.execute(*args, cwd=install_dir)
# gpu-burn reports on stdout
LOG.debug(out)
except (processutils.ProcessExecutionError, OSError) as e:
error_msg = 'gpu-burn failed with error %s' % e
LOG.error(error_msg)
raise errors.CommandExecutionError(error_msg)
gpu_data = [i for i in out.splitlines() if i.startswith('ID')]
gpu_count = len(gpu_data)
if gpu_count != count:
error_msg = ("gpu-burn failed to find the correct number of gpus. "
"%s found but %s expected." % (gpu_count, count))
LOG.error(error_msg)
raise errors.CleaningError(error_msg)
def _gpu_burn_run(install_dir, memory, timeout=86400):
"""Burn-in the GPU with gpu-burn
Run a GPU burn-in job for a configurable amount of time.
:param install_dir: The location where gpu-burn has been installed.
:param memory: Use N% or X MB of the available GPU memory.
:param timeout: Timeout in seconds (default: 86400).
:raises: CommandExecutionError if the execution of gpu-burn fails.
"""
args = ['./gpu_burn', '-m', memory, timeout]
LOG.debug('Burn-in gpu command: %s', args)
try:
out, _ = utils.execute(*args, cwd=install_dir)
# gpu-burn reports on stdout
LOG.debug(out)
except (processutils.ProcessExecutionError, OSError) as e:
error_msg = 'gpu-burn failed with error %s' % e
LOG.error(error_msg)
raise errors.CommandExecutionError(error_msg)
def gpu_burn(node):
"""Burn-in and check correct count of GPUs using gpu-burn
Check that the expected number of GPUs are available on the node
and run a GPU burn-in job for a configurable amount of time.
:param node: Ironic node object
"""
info = node.get('driver_info', {})
install_dir = info.get('agent_burnin_gpu_install_dir', '/opt/gpu-burn')
timeout = info.get('agent_burnin_gpu_timeout', 86400)
memory = info.get('agent_burnin_gpu_memory', '95%')
count = info.get('agent_burnin_gpu_count', 0)
# Only check count if an expected number of GPUs has been configured
if count > 0:
_gpu_burn_check_count(install_dir, count)
else:
LOG.debug("Burn-in gpu skipping expected number of GPUs check as "
"'agent_burnin_gpu_count' set to 0")
_gpu_burn_run(install_dir, memory, timeout)

22
ironic_python_agent/hardware.py

@ -2063,6 +2063,14 @@ class GenericHardwareManager(HardwareManager):
"""
burnin.stress_ng_cpu(node)
def burnin_gpu(self, node, ports):
"""Burn-in the GPU
:param node: Ironic node object
:param ports: list of Ironic port objects
"""
burnin.gpu_burn(node)
def burnin_disk(self, node, ports):
"""Burn-in the disk
@ -2656,6 +2664,13 @@ class GenericHardwareManager(HardwareManager):
'reboot_requested': False,
'abortable': True
},
{
'step': 'burnin_gpu',
'priority': 0,
'interface': 'deploy',
'reboot_requested': False,
'abortable': True
},
{
'step': 'burnin_disk',
'priority': 0,
@ -2752,6 +2767,13 @@ class GenericHardwareManager(HardwareManager):
'reboot_requested': False,
'abortable': True
},
{
'step': 'burnin_gpu',
'priority': 0,
'interface': 'deploy',
'reboot_requested': False,
'abortable': True
},
# NOTE(TheJulia): Burnin disk is explicitly not carried in this
# list because it would be destructive to data on a disk.
# If someone needs to do that, the machine should be

46
ironic_python_agent/tests/unit/test_burnin.py

@ -11,6 +11,7 @@
# under the License.
from unittest import mock
from unittest.mock import call
from oslo_concurrency import processutils
from tooz import coordination
@ -94,7 +95,6 @@ class TestBurnin(base.IronicAgentTest):
burnin.stress_ng_cpu, node)
def test_stress_ng_vm_default(self, mock_execute):
node = {'driver_info': {}}
mock_execute.return_value = (['out', 'err'])
@ -102,8 +102,8 @@ class TestBurnin(base.IronicAgentTest):
mock_execute.assert_called_once_with(
'stress-ng', '--vm', 0, '--vm-bytes', '98%',
'--timeout', 86400, '--metrics-brief')
'stress-ng', '--vm', 0, '--timeout', 86400, '--vm-bytes', '98%',
'--metrics-brief')
def test_stress_ng_vm_non_default(self, mock_execute):
@ -117,9 +117,8 @@ class TestBurnin(base.IronicAgentTest):
burnin.stress_ng_vm(node)
mock_execute.assert_called_once_with(
'stress-ng', '--vm', 2, '--vm-bytes', '25%',
'--timeout', 120, '--metrics-brief',
'--log-file', '/var/log/burnin.vm')
'stress-ng', '--vm', 2, '--timeout', 120, '--vm-bytes', '25%',
'--metrics-brief', '--log-file', '/var/log/burnin.vm')
def test_stress_ng_vm_no_stress_ng(self, mock_execute):
@ -515,3 +514,38 @@ class TestBurnin(base.IronicAgentTest):
# get_members is called initially, then every second until the
# other node appears
self.assertEqual(3, mock_coordinator.get_members.call_count)
def test_gpu_burn_default(self, mock_execute):
node = {'driver_info': {}}
mock_execute.return_value = (['out', 'err'])
expected_calls = [call('./gpu_burn', '-l', cwd='/opt/gpu-burn'),
call('./gpu_burn', '-m', '95%', 86400,
cwd='/opt/gpu-burn')]
burnin.gpu_burn(node)
mock_execute.assert_has_calls(expected_calls, any_order=False)
@mock.patch('ironic_python_agent.burnin.len', return_value=2, autospec=True)
def test_gpu_burn_non_default(self, mock_gpu_burn_check_count, mock_execute):
node = {'driver_info': {
'agent_burnin_gpu_count': 2,
'agent_burnin_gpu_timeout': 3600}}
mock_execute.return_value = (['out', 'err'])
expected_calls = [call('./gpu_burn', '-l', cwd='/opt/gpu-burn'),
call('./gpu_burn', '-m', '95%', 3600,
cwd='/opt/gpu-burn')]
burnin.gpu_burn(node)
mock_execute.assert_has_calls(expected_calls, any_order=False)
def test_gpu_burn_no_package(self, mock_execute):
node = {'driver_info': {}}
mock_execute.side_effect = processutils.ProcessExecutionError()
self.assertRaises(errors.CommandExecutionError,
burnin.gpu_burn, node)

7
ironic_python_agent/tests/unit/test_hardware.py

@ -253,6 +253,13 @@ class TestGenericHardwareManager(base.IronicAgentTest):
'reboot_requested': False,
'abortable': True
},
{
'step': 'burnin_gpu',
'priority': 0,
'interface': 'deploy',
'reboot_requested': False,
'abortable': True
},
{
'step': 'burnin_disk',
'priority': 0,

3
releasenotes/notes/add-support-for-burnin-gpu-76c8c267529a18bd.yaml Normal file

@ -0,0 +1,3 @@
features:
- |
Add support for GPU burn-in testing using gpu-burn.