def check_from_target(cls, target):
for domain in target.cpufreq.iter_domains():
if "schedutil" not in target.cpufreq.list_governors(domain[0]):
raise CannotCreateError(
"Can't set schedutil governor for domain {}".format(domain))
if 'nrg-model' not in target.plat_info:
raise CannotCreateError("Energy model not available")
def check_from_target(cls, target):
super().check_from_target(target)
try:
target.plat_info["cpu-capacities"]
except KeyError as e:
raise CannotCreateError(str(e))
def check_from_target(cls, target):
super().check_from_target(target)
cpus = cls.get_migration_cpus(target.plat_info)
if not len(cpus) == cls.NR_REQUIRED_CPUS:
raise CannotCreateError(
"This workload requires {} CPUs of identical capacity".format(
cls.NR_REQUIRED_CPUS))
def check_from_target(cls, target):
super().check_from_target(target)
try:
target.plat_info["cpu-capacities"]['rtapp']
except KeyError as e:
raise CannotCreateError(str(e))
# Check that there are enough CPUs of the same capacity
cls.get_migration_cpus(target.plat_info)
def get_migration_cpus(cls, plat_info):
"""
:returns: N CPUs of same capacity, with N set by :meth:`get_nr_required_cpu`.
"""
# Iterate over descending CPU capacity groups
nr_required_cpu = cls.get_nr_required_cpu(plat_info)
for cpus in reversed(plat_info["capacity-classes"]):
if len(cpus) >= nr_required_cpu:
return cpus[:nr_required_cpu]
raise CannotCreateError(
"This workload requires {} CPUs of identical capacity".format(
nr_required_cpu))
def exp_power(row):
task_utils = row.to_dict()
try:
expected_utils = nrg_model.get_optimal_placements(task_utils, capacity_margin_pct)[0]
except EnergyModelCapacityError:
raise CannotCreateError(
'The workload will result in overutilized status for all possible task placement, making it unsuitable to test EAS on this platform'
)
power = nrg_model.estimate_from_cpu_util(expected_utils)
columns = list(power.keys())
# Assemble a dataframe to plot the expected utilization
data.append(expected_utils)
index.append(row.name)
return pd.Series([power[c] for c in columns], index=columns)
def get_migration_cpus(cls, plat_info):
"""
:returns: N CPUs of same capacity, with N set by :meth:`get_nr_required_cpu`.
"""
# Iterate over descending CPU capacity groups
nr_required_cpu = cls.get_nr_required_cpu(plat_info)
cpu_classes = plat_info["capacity-classes"]
# If the CPU capacities are writeable, it's better to give priority to
# LITTLE cores as they will be less prone to thermal capping.
# Otherwise, it's better to pick big cores as they will not be affected
# by CPU invariance issues.
if not plat_info['cpu-capacities']['writeable']:
cpu_classes = reversed(cpu_classes)
for cpus in cpu_classes:
if len(cpus) >= nr_required_cpu:
return cpus[:nr_required_cpu]
raise CannotCreateError(
f"This workload requires {nr_required_cpu} CPUs of identical capacity")
def check_from_target(cls, target):
if target.number_of_nodes < 2:
raise CannotCreateError(
"Target doesn't have at least two NUMA nodes")
def get_simulated_pelt(self, task, signal_name):
"""
Simulate a PELT signal for a given task.
:param task: task to look for in the trace.
:type task: int or str or tuple(int, str)
:param signal_name: Name of the PELT signal to simulate.
:type signal_name: str
:return: A :class:`pandas.DataFrame` with a ``simulated`` column
containing the simulated signal, along with the column of the
signal as found in the trace.
"""
logger = self.get_logger()
trace = self.trace
task = trace.get_task_id(task)
cpus = trace.analysis.tasks.cpus_of_tasks([task])
df_activation = trace.analysis.tasks.df_task_activation(
task,
# Util only takes into account times where the task is actually
# executing
preempted_value=0,
)
df = trace.analysis.load_tracking.df_tasks_signal(signal_name)
df = df_filter_task_ids(df, [task])
df = df.copy(deep=False)
# Ignore the first activation, as its signals are incorrect
df_activation = df_activation.iloc[2:]
# Make sure the activation df does not start before the dataframe of
# signal values, otherwise we cannot provide a sensible init value
df_activation = df_activation[df.index[0]:]
# Get the initial signal value matching the first activation we will care about
init_iloc = df.index.get_loc(df_activation.index[0], method='ffill')
init = df[signal_name].iloc[init_iloc]
try:
# PELT clock in nanoseconds
clock = df['update_time'] * 1e-9
except KeyError:
if any(self.plat_info['cpu-capacities']['rtapp'][cpu] != UTIL_SCALE
for phase in self.wlgen_task.phases for cpu in phase.cpus):
raise CannotCreateError(
'PELT time scaling can only be simulated when the PELT clock is available from the trace'
)
logger.warning(
'PELT clock is not available, ftrace timestamp will be used at the expense of accuracy'
)
clock = None
df['simulated'] = simulate_pelt(df_activation['active'],
index=df.index,
init=init,
clock=clock)
# Since load is now CPU invariant in recent kernel versions, we don't
# rescale it back. To match the old behavior, that line is
# needed:
# df['simulated'] /= self.plat_info['cpu-capacities']['rtapp'][cpu] / UTIL_SCALE
kernel_version = self.plat_info['kernel']['version']
if (signal_name == 'load' and kernel_version.parts[:2] < (5, 1)):
logger().warning(
f'Load signal is assumed to be CPU invariant, which is true for recent mainline kernels, but may be wrong for {kernel_version}'
)
df['error'] = df[signal_name] - df['simulated']
df = df.dropna()
return df
def check_from_target(cls, target):
if not cls._has_asym_cpucapacity(target):
raise CannotCreateError(
"Target doesn't have asymmetric CPU capacities")
def check_from_target(cls, target):
if len(target.plat_info["capacity-classes"]) < 2:
raise CannotCreateError(
'Cannot test migration on single capacity group')
def check_from_target(cls, target):
if not cls._has_asym_cpucapacity(target):
raise CannotCreateError(
"Target doesn't have SD_ASYM_CPUCAPACITY on any sched_domain")