def df_peripheral_clock_effective_rate(self, clk_name):
# Note: the kernel still defines a "clock_*" variant for each of these,
# but it's not actually used anywhere in the code. The new "clk_*"
# events are the ones we are interested about.
rate_df = self.trace.df_event('clk_set_rate')
enable_df = self.trace.df_event('clk_enable').copy()
disable_df = self.trace.df_event('clk_disable').copy()
# Add 'state' for enable and disable events
enable_df['state'] = 1
disable_df['state'] = 0
freq = rate_df[rate_df['name'] == clk_name]
enables = enable_df[enable_df['name'] == clk_name]
disables = disable_df[disable_df['name'] == clk_name]
freq = df_merge((freq, enables, disables)).ffill()
freq['start'] = freq.index
df_add_delta(freq,
col='len',
src_col='start',
window=self.trace.window,
inplace=True)
freq['effective_rate'] = np.where(freq['state'] == 0, 0, freq['rate'])
return freq
def _get_callgraph(self, tag_df=None, thread_root_functions=None):
entry_df = self.df_funcgraph(event='entry').copy(deep=False)
entry_df['event'] = _CallGraph._EVENT.ENTRY
exit_df = self.df_funcgraph(event='exit').copy(deep=False)
exit_df['event'] = _CallGraph._EVENT.EXIT
# Attempt to get the CPU capacity signal to normalize the results
capacity_cols = ['__cpu', 'event', 'capacity']
try:
capacity_df = self.trace.analysis.load_tracking.df_cpus_signal(
'capacity')
except MissingTraceEventError:
capacity_df = pd.DataFrame(columns=capacity_cols)
else:
capacity_df = capacity_df.copy(deep=False)
capacity_df['__cpu'] = capacity_df['cpu']
capacity_df['event'] = _CallGraph._EVENT.SET_CAPACITY
capacity_df = capacity_df[capacity_cols]
# Set a reasonable initial capacity
try:
orig_capacities = self.trace.plat_info['cpu-capacities']['orig']
except KeyError:
pass
else:
orig_capacities_df = pd.DataFrame.from_records(
((-1 * cpu, cpu, _CallGraph._EVENT.SET_CAPACITY, cap)
for cpu, cap in orig_capacities.items()),
columns=['Time', '__cpu', 'event', 'capacity'],
index='Time',
)
capacity_df = pd.concat((orig_capacities_df, capacity_df))
to_merge = [entry_df, exit_df, capacity_df]
if tag_df is not None:
cpu = tag_df['__cpu']
tag_df = tag_df.drop(columns=['__cpu'])
tag_df = pd.DataFrame(
dict(
tags=tag_df.apply(pd.Series.to_dict, axis=1),
__cpu=cpu,
))
tag_df['event'] = _CallGraph._EVENT.SET_TAG
to_merge.append(tag_df)
df = df_merge(to_merge)
return _CallGraph.from_df(df,
thread_root_functions=thread_root_functions)
def compare_with_traces(self, others, normalize=True, **kwargs):
"""
Compare the :class:`~lisa.trace.Trace` it's called on with the other
traces passed as ``others``. The reference is the trace it's called on.
:returns: a :class:`lisa.stats.Stats` object just like
:meth:`profile_stats`.
:param others: List of traces to compare against.
:type others: list(lisa.trace.Trace)
:Variable keyword arguments: Forwarded to :meth:`profile_stats`.
"""
ref = self.trace
traces = [ref] + list(others)
paths = [trace.trace_path for trace in traces]
common_prefix_len = len(os.path.commonprefix(paths))
common_suffix_len = len(
os.path.commonprefix(list(map(lambda x: str(reversed(x)), paths))))
def get_name(trace):
name = trace.trace_path[common_prefix_len:common_suffix_len]
if not name:
if trace is ref:
name = 'ref'
else:
name = str(traces.index(trace))
return name
def get_df(trace):
df = self.df_calls(normalize=normalize)
df = df.copy(deep=False)
df['trace'] = get_name(trace)
return df
df = df_merge(map(get_df, traces))
ref_group = {'trace': get_name(ref)}
return self._profile_stats_from_df(df, ref_group=ref_group, **kwargs)
def df_ramp_boost(self):
"""
Return a dataframe with schedutil-related signals, sampled at the
frequency decisions timestamps for the CPU this bundle was executed on.
.. note:: The computed columns only take into account the CPU the test
was executing on. It currently does not handle multi-task workloads.
"""
trace = self.trace
cpu = self.cpu
task = self.rtapp_task_ids[0]
# schedutil_df also has a 'util' column that would conflict
schedutil_df = trace.df_event('schedutil_em')[['cpu', 'cost_margin', 'base_freq']]
schedutil_df = schedutil_df.copy()
schedutil_df['from_schedutil'] = True
def compute_base_cost(row):
freq = row['base_freq']
cpu = row['cpu']
em = self.plat_info['nrg-model']
active_states = em.cpu_nodes[cpu].active_states
freqs = sorted(active_states.keys())
max_freq = max(freqs)
def cost(freq):
higher_freqs = list(itertools.dropwhile(lambda f: f < freq, freqs))
freq = freqs[-1] if not higher_freqs else higher_freqs[0]
active_state = active_states[freq]
return active_state.power * max_freq / freq
max_cost = max(
cost(freq)
for freq in active_states.keys()
)
return cost(freq) / max_cost * 100
schedutil_df['base_cost'] = schedutil_df.apply(compute_base_cost, axis=1)
task_active = trace.analysis.tasks.df_task_states(task)['curr_state']
task_active = task_active.apply(lambda state: int(state == TaskState.TASK_ACTIVE))
task_active = task_active.reindex(schedutil_df.index, method='ffill')
# Assume task active == CPU active, since there is only one task
assert len(self.rtapp_task_ids) == 1
cpu_active_df = pd.DataFrame({'cpu_active': task_active})
cpu_active_df['cpu'] = cpu
cpu_active_df.dropna(inplace=True)
df_list = [
schedutil_df,
trace.analysis.load_tracking.df_cpus_signal('util'),
trace.analysis.load_tracking.df_cpus_signal('enqueued'),
cpu_active_df,
]
df = df_merge(df_list, filter_columns={'cpu': cpu})
df['from_schedutil'].fillna(value=False, inplace=True)
df.ffill(inplace=True)
df.dropna(inplace=True)
# Reconstitute how schedutil sees signals by subsampling the
# "master" dataframe, so we can look at signals coming from other
# dataframes
df = df[df['from_schedutil'] == True] # pylint: disable=singleton-comparison
df.drop(columns=['from_schedutil'], inplace=True)
# If there are some NaN at the beginning, just drop some data rather
# than using fake data
df.dropna(inplace=True)
boost_points = (
# util_est_enqueued is the same as last freq update
(df['enqueued'].diff() == 0) &
# util_avg is increasing
(df['util'].diff() >= 0) &
# util_avg > util_est_enqueued
(df['util'] > df['enqueued']) &
# CPU is not idle
(df['cpu_active'])
)
df['boost_points'] = boost_points
df['expected_cost_margin'] = (df['util'] - df['enqueued']).where(
cond=boost_points,
other=0,
)
# cost_margin values range from 0 to 1024
ENERGY_SCALE = 1024
for col in ('expected_cost_margin', 'cost_margin'):
df[col] *= 100 / ENERGY_SCALE
df['allowed_cost'] = df['base_cost'] + df['cost_margin']
# We cannot know if the first row is supposed to be boosted or not
# because we lack history, so we just drop it
return df.iloc[1:]
def test_hotplug_rollback(self) -> ResultBundle:
"""
Test that the hotplug can rollback to its previous state after a
failure. All possible steps, up/down combinations will be tested. For
each combination, also verify that the hotplug is going through all the
steps it is supposed to.
"""
df = df_merge([
self.trace.df_event('userspace@hotplug_rollback'),
self.trace.df_event('cpuhp_enter')
])
# Keep only the states delimited by _mark_trace()
df['test'].ffill(inplace=True)
df = df[df['test'] == 1]
df.drop(columns='test', inplace=True)
df['up'].ffill(inplace=True)
df['up'] = df['up'].astype(bool)
# Read the expected states from full hot(un)plug
df['expected'].ffill(inplace=True)
df['expected'] = df['expected'].astype(bool)
expected_down = self._get_expected_states(df, up=False)
expected_up = self._get_expected_states(df, up=True)
df = df[~df['expected']]
df.drop(columns='expected', inplace=True)
def _get_expected_rollback(up, failing_state):
return list(
filter(
partial(
operator.gt if up else operator.lt,
failing_state,
),
chain(expected_up, expected_down) if up else
chain(expected_down, expected_up)
)
)
def _verify_rollback(df):
failing_state = df['failing_state'].iloc[0]
up = df['up'].iloc[0]
expected = _get_expected_rollback(up, failing_state)
return pd.DataFrame(data={
'failing_state': df['failing_state'],
'up': up,
'result': df['idx'].tolist() == expected
})
df['failing_state'].ffill(inplace=True)
df.dropna(inplace=True)
df = df.groupby(['up', 'failing_state'],
observed=True).apply(_verify_rollback)
df.drop_duplicates(inplace=True)
res = ResultBundle.from_bool(df['result'].all())
res.add_metric('Failed rollback states',
df[~df['result']]['failing_state'].tolist())
return res
def df_ramp_boost(self):
"""
Return a dataframe with schedutil-related signals, sampled at the
frequency decisions timestamps for the CPU this bundle was executed on.
.. note:: The computed columns only take into account the CPU the test
was executing on. It currently does not handle multi-task workloads.
"""
trace = self.trace
cpu = self.cpu
# schedutil_df also has a 'util' column that would conflict
schedutil_df = trace.df_events('schedutil_em')[['cpu', 'cost_margin']]
schedutil_df['from_schedutil'] = True
df_list = [
schedutil_df,
trace.analysis.load_tracking.df_cpus_signal('util'),
trace.analysis.load_tracking.df_cpus_signal('util_est_enqueued'),
]
df = df_merge(df_list, filter_columns={'cpu': cpu})
df['from_schedutil'].fillna(value=False, inplace=True)
df.ffill(inplace=True)
df.dropna(inplace=True)
# Reconstitute how schedutil sees signals by subsampling the
# "master" dataframe, so we can look at signals coming from other
# dataframes
df = df[df['from_schedutil'] == True]
df.drop(columns=['from_schedutil'], inplace=True)
# If there are some NaN at the beginning, just drop some data rather
# than using fake data
df.dropna(inplace=True)
boost_points = (
# util_est_enqueued is the same as last freq update
(df['util_est_enqueued'].diff() == 0) &
# util_avg is increasing
(df['util'].diff() >= 0) &
# util_avg > util_est_enqueued
(df['util'] > df['util_est_enqueued']))
df['boost_points'] = boost_points
df['expected_cost_margin'] = (df['util'] -
df['util_est_enqueued']).where(
cond=boost_points,
other=0,
)
# cost_margin values range from 0 to 1024
ENERGY_SCALE = 1024
for col in ('expected_cost_margin', 'cost_margin'):
df[col] *= 100 / ENERGY_SCALE
# We cannot know if the first row is supposed to be boosted or not
# because we lack history, so we just drop it
return df.iloc[1:]
