mirror of
https://github.com/ARM-software/workload-automation.git
synced 2024-10-06 19:01:15 +01:00
851 lines
41 KiB
Python
851 lines
41 KiB
Python
# Copyright 2015 ARM Limited
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
#
|
|
|
|
|
|
#pylint: disable=attribute-defined-outside-init,access-member-before-definition,redefined-outer-name
|
|
from __future__ import division
|
|
import os
|
|
import math
|
|
import time
|
|
from tempfile import mktemp
|
|
from base64 import b64encode
|
|
from collections import Counter, namedtuple
|
|
|
|
try:
|
|
import jinja2
|
|
import pandas as pd
|
|
import matplotlib
|
|
matplotlib.use('AGG')
|
|
import matplotlib.pyplot as plt
|
|
import numpy as np
|
|
low_filter = np.vectorize(lambda x: x > 0 and x or 0) # pylint: disable=no-member
|
|
import_error = None
|
|
except ImportError as e:
|
|
import_error = e
|
|
jinja2 = None
|
|
pd = None
|
|
plt = None
|
|
np = None
|
|
low_filter = None
|
|
|
|
from wlauto import Instrument, Parameter, File
|
|
from wlauto.exceptions import ConfigError, InstrumentError, DeviceError
|
|
from wlauto.instrumentation import instrument_is_installed
|
|
from wlauto.utils.types import caseless_string, list_or_caseless_string, list_of_ints
|
|
from wlauto.utils.misc import list_to_mask
|
|
|
|
FREQ_TABLE_FILE = 'frequency_power_perf_data.csv'
|
|
CPUS_TABLE_FILE = 'projected_cap_power.csv'
|
|
MEASURED_CPUS_TABLE_FILE = 'measured_cap_power.csv'
|
|
IDLE_TABLE_FILE = 'idle_power_perf_data.csv'
|
|
REPORT_TEMPLATE_FILE = 'report.template'
|
|
EM_TEMPLATE_FILE = 'em.template'
|
|
|
|
IdlePowerState = namedtuple('IdlePowerState', ['power'])
|
|
CapPowerState = namedtuple('CapPowerState', ['cap', 'power'])
|
|
|
|
|
|
class EnergyModel(object):
|
|
|
|
def __init__(self):
|
|
self.big_cluster_idle_states = []
|
|
self.little_cluster_idle_states = []
|
|
self.big_cluster_cap_states = []
|
|
self.little_cluster_cap_states = []
|
|
self.big_core_idle_states = []
|
|
self.little_core_idle_states = []
|
|
self.big_core_cap_states = []
|
|
self.little_core_cap_states = []
|
|
|
|
def add_cap_entry(self, cluster, perf, clust_pow, core_pow):
|
|
if cluster == 'big':
|
|
self.big_cluster_cap_states.append(CapPowerState(perf, clust_pow))
|
|
self.big_core_cap_states.append(CapPowerState(perf, core_pow))
|
|
elif cluster == 'little':
|
|
self.little_cluster_cap_states.append(CapPowerState(perf, clust_pow))
|
|
self.little_core_cap_states.append(CapPowerState(perf, core_pow))
|
|
else:
|
|
raise ValueError('Unexpected cluster: {}'.format(cluster))
|
|
|
|
def add_cluster_idle(self, cluster, values):
|
|
for value in values:
|
|
if cluster == 'big':
|
|
self.big_cluster_idle_states.append(IdlePowerState(value))
|
|
elif cluster == 'little':
|
|
self.little_cluster_idle_states.append(IdlePowerState(value))
|
|
else:
|
|
raise ValueError('Unexpected cluster: {}'.format(cluster))
|
|
|
|
def add_core_idle(self, cluster, values):
|
|
for value in values:
|
|
if cluster == 'big':
|
|
self.big_core_idle_states.append(IdlePowerState(value))
|
|
elif cluster == 'little':
|
|
self.little_core_idle_states.append(IdlePowerState(value))
|
|
else:
|
|
raise ValueError('Unexpected cluster: {}'.format(cluster))
|
|
|
|
|
|
class PowerPerformanceAnalysis(object):
|
|
|
|
def __init__(self, data):
|
|
self.summary = {}
|
|
big_freqs = data[data.cluster == 'big'].frequency.unique()
|
|
little_freqs = data[data.cluster == 'little'].frequency.unique()
|
|
self.summary['frequency'] = max(set(big_freqs).intersection(set(little_freqs)))
|
|
|
|
big_sc = data[(data.cluster == 'big') &
|
|
(data.frequency == self.summary['frequency']) &
|
|
(data.cpus == 1)]
|
|
little_sc = data[(data.cluster == 'little') &
|
|
(data.frequency == self.summary['frequency']) &
|
|
(data.cpus == 1)]
|
|
self.summary['performance_ratio'] = big_sc.performance.item() / little_sc.performance.item()
|
|
self.summary['power_ratio'] = big_sc.power.item() / little_sc.power.item()
|
|
self.summary['max_performance'] = data[data.cpus == 1].performance.max()
|
|
self.summary['max_power'] = data[data.cpus == 1].power.max()
|
|
|
|
|
|
def build_energy_model(freq_power_table, cpus_power, idle_power, first_cluster_idle_state):
|
|
# pylint: disable=too-many-locals
|
|
em = EnergyModel()
|
|
idle_power_sc = idle_power[idle_power.cpus == 1]
|
|
perf_data = get_normalized_single_core_data(freq_power_table)
|
|
|
|
for cluster in ['little', 'big']:
|
|
cluster_cpus_power = cpus_power[cluster].dropna()
|
|
cluster_power = cluster_cpus_power['cluster'].apply(int)
|
|
core_power = (cluster_cpus_power['1'] - cluster_power).apply(int)
|
|
performance = (perf_data[perf_data.cluster == cluster].performance_norm * 1024 / 100).apply(int)
|
|
for perf, clust_pow, core_pow in zip(performance, cluster_power, core_power):
|
|
em.add_cap_entry(cluster, perf, clust_pow, core_pow)
|
|
|
|
all_idle_power = idle_power_sc[idle_power_sc.cluster == cluster].power.values
|
|
# CORE idle states
|
|
# We want the delta of each state w.r.t. the power
|
|
# consumption of the shallowest one at this level (core_ref)
|
|
idle_core_power = low_filter(all_idle_power[:first_cluster_idle_state] -
|
|
all_idle_power[first_cluster_idle_state - 1])
|
|
# CLUSTER idle states
|
|
# We want the absolute value of each idle state
|
|
idle_cluster_power = low_filter(all_idle_power[first_cluster_idle_state - 1:])
|
|
em.add_cluster_idle(cluster, idle_cluster_power)
|
|
em.add_core_idle(cluster, idle_core_power)
|
|
|
|
return em
|
|
|
|
|
|
def generate_em_c_file(em, big_core, little_core, em_template_file, outfile):
|
|
with open(em_template_file) as fh:
|
|
em_template = jinja2.Template(fh.read())
|
|
em_text = em_template.render(
|
|
big_core=big_core,
|
|
little_core=little_core,
|
|
em=em,
|
|
)
|
|
with open(outfile, 'w') as wfh:
|
|
wfh.write(em_text)
|
|
return em_text
|
|
|
|
|
|
def generate_report(freq_power_table, measured_cpus_table, cpus_table, idle_power_table, # pylint: disable=unused-argument
|
|
report_template_file, device_name, em_text, outfile):
|
|
# pylint: disable=too-many-locals
|
|
cap_power_analysis = PowerPerformanceAnalysis(freq_power_table)
|
|
single_core_norm = get_normalized_single_core_data(freq_power_table)
|
|
cap_power_plot = get_cap_power_plot(single_core_norm)
|
|
idle_power_plot = get_idle_power_plot(idle_power_table)
|
|
|
|
fig, axes = plt.subplots(1, 2)
|
|
fig.set_size_inches(16, 8)
|
|
for i, cluster in enumerate(reversed(cpus_table.columns.levels[0])):
|
|
projected = cpus_table[cluster].dropna(subset=['1'])
|
|
plot_cpus_table(projected, axes[i], cluster)
|
|
cpus_plot_data = get_figure_data(fig)
|
|
|
|
with open(report_template_file) as fh:
|
|
report_template = jinja2.Template(fh.read())
|
|
html = report_template.render(
|
|
device_name=device_name,
|
|
freq_power_table=freq_power_table.set_index(['cluster', 'cpus', 'frequency']).to_html(),
|
|
cap_power_analysis=cap_power_analysis,
|
|
cap_power_plot=get_figure_data(cap_power_plot),
|
|
idle_power_table=idle_power_table.set_index(['cluster', 'cpus', 'state']).to_html(),
|
|
idle_power_plot=get_figure_data(idle_power_plot),
|
|
cpus_table=cpus_table.to_html(),
|
|
cpus_plot=cpus_plot_data,
|
|
em_text=em_text,
|
|
)
|
|
with open(outfile, 'w') as wfh:
|
|
wfh.write(html)
|
|
return html
|
|
|
|
|
|
def wa_result_to_power_perf_table(df, performance_metric, index):
|
|
table = df.pivot_table(index=index + ['iteration'],
|
|
columns='metric', values='value').reset_index()
|
|
result_mean = table.groupby(index).mean()
|
|
result_std = table.groupby(index).std()
|
|
result_std.columns = [c + ' std' for c in result_std.columns]
|
|
result_count = table.groupby(index).count()
|
|
result_count.columns = [c + ' count' for c in result_count.columns]
|
|
count_sqrt = result_count.apply(lambda x: x.apply(math.sqrt))
|
|
count_sqrt.columns = result_std.columns # match column names for division
|
|
result_error = 1.96 * result_std / count_sqrt # 1.96 == 95% confidence interval
|
|
result_error.columns = [c + ' error' for c in result_mean.columns]
|
|
|
|
result = pd.concat([result_mean, result_std, result_count, result_error], axis=1)
|
|
del result['iteration']
|
|
del result['iteration std']
|
|
del result['iteration count']
|
|
del result['iteration error']
|
|
|
|
updated_columns = []
|
|
for column in result.columns:
|
|
if column == performance_metric:
|
|
updated_columns.append('performance')
|
|
elif column == performance_metric + ' std':
|
|
updated_columns.append('performance_std')
|
|
elif column == performance_metric + ' error':
|
|
updated_columns.append('performance_error')
|
|
else:
|
|
updated_columns.append(column.replace(' ', '_'))
|
|
result.columns = updated_columns
|
|
result = result[sorted(result.columns)]
|
|
result.reset_index(inplace=True)
|
|
|
|
return result
|
|
|
|
|
|
def get_figure_data(fig, fmt='png'):
|
|
tmp = mktemp()
|
|
fig.savefig(tmp, format=fmt, bbox_inches='tight')
|
|
with open(tmp, 'rb') as fh:
|
|
image_data = b64encode(fh.read())
|
|
os.remove(tmp)
|
|
return image_data
|
|
|
|
|
|
def get_normalized_single_core_data(data):
|
|
finite_power = np.isfinite(data.power) # pylint: disable=no-member
|
|
finite_perf = np.isfinite(data.performance) # pylint: disable=no-member
|
|
data_single_core = data[(data.cpus == 1) & finite_perf & finite_power].copy()
|
|
data_single_core['performance_norm'] = (data_single_core.performance /
|
|
data_single_core.performance.max() * 100).apply(int)
|
|
data_single_core['power_norm'] = (data_single_core.power /
|
|
data_single_core.power.max() * 100).apply(int)
|
|
return data_single_core
|
|
|
|
|
|
def get_cap_power_plot(data_single_core):
|
|
big_single_core = data_single_core[(data_single_core.cluster == 'big') &
|
|
(data_single_core.cpus == 1)]
|
|
little_single_core = data_single_core[(data_single_core.cluster == 'little') &
|
|
(data_single_core.cpus == 1)]
|
|
|
|
fig, axes = plt.subplots(1, 1, figsize=(12, 8))
|
|
axes.plot(big_single_core.performance_norm,
|
|
big_single_core.power_norm,
|
|
marker='o')
|
|
axes.plot(little_single_core.performance_norm,
|
|
little_single_core.power_norm,
|
|
marker='o')
|
|
axes.set_xlim(0, 105)
|
|
axes.set_ylim(0, 105)
|
|
axes.set_xlabel('Performance (Normalized)')
|
|
axes.set_ylabel('Power (Normalized)')
|
|
axes.grid()
|
|
axes.legend(['big cluster', 'little cluster'], loc=0)
|
|
return fig
|
|
|
|
|
|
def get_idle_power_plot(df):
|
|
fig, axes = plt.subplots(1, 2, figsize=(15, 7))
|
|
for cluster, ax in zip(['little', 'big'], axes):
|
|
data = df[df.cluster == cluster].pivot_table(index=['state'], columns='cpus', values='power')
|
|
err = df[df.cluster == cluster].pivot_table(index=['state'], columns='cpus', values='power_error')
|
|
data.plot(kind='bar', ax=ax, rot=30, yerr=err)
|
|
ax.set_title('{} cluster'.format(cluster))
|
|
ax.set_xlim(-1, len(data.columns) - 0.5)
|
|
ax.set_ylabel('Power (mW)')
|
|
return fig
|
|
|
|
|
|
def fit_polynomial(s, n):
|
|
# pylint: disable=no-member
|
|
coeffs = np.polyfit(s.index, s.values, n)
|
|
poly = np.poly1d(coeffs)
|
|
return poly(s.index)
|
|
|
|
|
|
def get_cpus_power_table(data, index, opps, leak_factors): # pylint: disable=too-many-locals
|
|
# pylint: disable=no-member
|
|
power_table = data[[index, 'cluster', 'cpus', 'power']].pivot_table(index=index,
|
|
columns=['cluster', 'cpus'],
|
|
values='power')
|
|
bs_power_table = pd.DataFrame(index=power_table.index, columns=power_table.columns)
|
|
for cluster in power_table.columns.levels[0]:
|
|
power_table[cluster, 0] = (power_table[cluster, 1] -
|
|
(power_table[cluster, 2] -
|
|
power_table[cluster, 1]))
|
|
bs_power_table.loc[power_table[cluster, 1].notnull(), (cluster, 1)] = fit_polynomial(power_table[cluster, 1].dropna(), 2)
|
|
bs_power_table.loc[power_table[cluster, 2].notnull(), (cluster, 2)] = fit_polynomial(power_table[cluster, 2].dropna(), 2)
|
|
|
|
if opps[cluster] is None:
|
|
bs_power_table.loc[bs_power_table[cluster, 1].notnull(), (cluster, 0)] = \
|
|
(2 * power_table[cluster, 1] - power_table[cluster, 2]).values
|
|
else:
|
|
voltages = opps[cluster].set_index('frequency').sort_index()
|
|
leakage = leak_factors[cluster] * 2 * voltages['voltage']**3 / 0.9**3
|
|
leakage_delta = leakage - leakage[leakage.index[0]]
|
|
bs_power_table.loc[:, (cluster, 0)] = \
|
|
(2 * bs_power_table[cluster, 1] + leakage_delta - bs_power_table[cluster, 2])
|
|
|
|
# re-order columns and rename colum '0' to 'cluster'
|
|
power_table = power_table[sorted(power_table.columns,
|
|
cmp=lambda x, y: cmp(y[0], x[0]) or cmp(x[1], y[1]))]
|
|
bs_power_table = bs_power_table[sorted(bs_power_table.columns,
|
|
cmp=lambda x, y: cmp(y[0], x[0]) or cmp(x[1], y[1]))]
|
|
old_levels = power_table.columns.levels
|
|
power_table.columns.set_levels([old_levels[0], list(map(str, old_levels[1])[:-1]) + ['cluster']],
|
|
inplace=True)
|
|
bs_power_table.columns.set_levels([old_levels[0], list(map(str, old_levels[1])[:-1]) + ['cluster']],
|
|
inplace=True)
|
|
return power_table, bs_power_table
|
|
|
|
|
|
def plot_cpus_table(projected, ax, cluster):
|
|
projected.T.plot(ax=ax, marker='o')
|
|
ax.set_title('{} cluster'.format(cluster))
|
|
ax.set_xticklabels(projected.columns)
|
|
ax.set_xticks(range(0, 5))
|
|
ax.set_xlim(-0.5, len(projected.columns) - 0.5)
|
|
ax.set_ylabel('Power (mW)')
|
|
ax.grid(True)
|
|
|
|
|
|
def opp_table(d):
|
|
if d is None:
|
|
return None
|
|
return pd.DataFrame(d.items(), columns=['frequency', 'voltage'])
|
|
|
|
|
|
class EnergyModelInstrument(Instrument):
|
|
|
|
name = 'energy_model'
|
|
desicription = """
|
|
Generates a power mode for the device based on specified workload.
|
|
|
|
This insturment will execute the workload specified by the agenda (currently, only ``sysbench`` is
|
|
supported) and will use the resulting performance and power measurments to generate a power mode for
|
|
the device.
|
|
|
|
This instrument requires certain features to be present in the kernel:
|
|
|
|
1. cgroups and cpusets must be enabled.
|
|
2. cpufreq and userspace governor must be enabled.
|
|
3. cpuidle must be enabled.
|
|
|
|
"""
|
|
|
|
parameters = [
|
|
Parameter('device_name', kind=caseless_string,
|
|
description="""The name of the device to be used in generating the model. If not specified,
|
|
``device.name`` will be used. """),
|
|
Parameter('big_core', kind=caseless_string,
|
|
description="""The name of the "big" core in the big.LITTLE system; must match
|
|
one of the values in ``device.core_names``. """),
|
|
Parameter('performance_metric', kind=caseless_string, mandatory=True,
|
|
description="""Metric to be used as the performance indicator."""),
|
|
Parameter('power_metric', kind=list_or_caseless_string,
|
|
description="""Metric to be used as the power indicator. The value may contain a
|
|
``{core}`` format specifier that will be replaced with names of big
|
|
and little cores to drive the name of the metric for that cluster.
|
|
Ether this or ``energy_metric`` must be specified but not both."""),
|
|
Parameter('energy_metric', kind=list_or_caseless_string,
|
|
description="""Metric to be used as the energy indicator. The value may contain a
|
|
``{core}`` format specifier that will be replaced with names of big
|
|
and little cores to drive the name of the metric for that cluster.
|
|
this metric will be used to derive power by deviding through by
|
|
execution time. Either this or ``power_metric`` must be specified, but
|
|
not both."""),
|
|
Parameter('power_scaling_factor', kind=float, default=1.0,
|
|
description="""Power model specfies power in milliWatts. This is a scaling factor that
|
|
power_metric values will be multiplied by to get milliWatts."""),
|
|
Parameter('big_frequencies', kind=list_of_ints,
|
|
description="""List of frequencies to be used for big cores. These frequencies must
|
|
be supported by the cores. If this is not specified, all available
|
|
frequencies for the core (as read from cpufreq) will be used."""),
|
|
Parameter('little_frequencies', kind=list_of_ints,
|
|
description="""List of frequencies to be used for little cores. These frequencies must
|
|
be supported by the cores. If this is not specified, all available
|
|
frequencies for the core (as read from cpufreq) will be used."""),
|
|
Parameter('idle_workload', kind=str, default='idle',
|
|
description="Workload to be used while measuring idle power."),
|
|
Parameter('idle_workload_params', kind=dict, default={},
|
|
description="Parameter to pass to the idle workload."),
|
|
Parameter('first_cluster_idle_state', kind=int, default=-1,
|
|
description='''The index of the first cluster idle state on the device. Previous states
|
|
are assumed to be core idles. The default is ``-1``, i.e. only the last
|
|
idle state is assumed to affect the entire cluster.'''),
|
|
Parameter('no_hotplug', kind=bool, default=False,
|
|
description='''This options allows running the instrument without hotpluging cores on and off.
|
|
Disabling hotplugging will most likely produce a less accurate power model.'''),
|
|
Parameter('num_of_freqs_to_thermal_adjust', kind=int, default=0,
|
|
description="""The number of frequencies begining from the highest, to be adjusted for
|
|
the thermal effect."""),
|
|
Parameter('big_opps', kind=opp_table,
|
|
description="""OPP table mapping frequency to voltage (kHz --> mV) for the big cluster."""),
|
|
Parameter('little_opps', kind=opp_table,
|
|
description="""OPP table mapping frequency to voltage (kHz --> mV) for the little cluster."""),
|
|
Parameter('big_leakage', kind=int, default=120,
|
|
description="""
|
|
Leakage factor for the big cluster (this is specific to a particular core implementation).
|
|
"""),
|
|
Parameter('little_leakage', kind=int, default=60,
|
|
description="""
|
|
Leakage factor for the little cluster (this is specific to a particular core implementation).
|
|
"""),
|
|
]
|
|
|
|
def validate(self):
|
|
if import_error:
|
|
message = 'energy_model instrument requires pandas, jinja2 and matplotlib Python packages to be installed; got: "{}"'
|
|
raise InstrumentError(message.format(import_error.message))
|
|
for capability in ['cgroups', 'cpuidle']:
|
|
if not self.device.has(capability):
|
|
message = 'The Device does not appear to support {}; does it have the right module installed?'
|
|
raise ConfigError(message.format(capability))
|
|
device_cores = set(self.device.core_names)
|
|
if (self.power_metric and self.energy_metric) or not (self.power_metric or self.energy_metric):
|
|
raise ConfigError('Either power_metric or energy_metric must be specified (but not both).')
|
|
if not device_cores:
|
|
raise ConfigError('The Device does not appear to have core_names configured.')
|
|
elif len(device_cores) != 2:
|
|
raise ConfigError('The Device does not appear to be a big.LITTLE device.')
|
|
if self.big_core and self.big_core not in self.device.core_names:
|
|
raise ConfigError('Specified big_core "{}" is in divice {}'.format(self.big_core, self.device.name))
|
|
if not self.big_core:
|
|
self.big_core = self.device.core_names[-1] # the last core is usually "big" in existing big.LITTLE devices
|
|
if not self.device_name:
|
|
self.device_name = self.device.name
|
|
if self.num_of_freqs_to_thermal_adjust and not instrument_is_installed('daq'):
|
|
self.logger.warn('Adjustment for thermal effect requires daq instrument. Disabling adjustment')
|
|
self.num_of_freqs_to_thermal_adjust = 0
|
|
|
|
def initialize(self, context):
|
|
self.number_of_cpus = {}
|
|
self.report_template_file = context.resolver.get(File(self, REPORT_TEMPLATE_FILE))
|
|
self.em_template_file = context.resolver.get(File(self, EM_TEMPLATE_FILE))
|
|
self.little_core = (set(self.device.core_names) - set([self.big_core])).pop()
|
|
self.perform_runtime_validation()
|
|
self.enable_all_cores()
|
|
self.configure_clusters()
|
|
self.discover_idle_states()
|
|
self.disable_thermal_management()
|
|
self.initialize_job_queue(context)
|
|
self.initialize_result_tracking()
|
|
|
|
def setup(self, context):
|
|
if not context.spec.label.startswith('idle_'):
|
|
return
|
|
for idle_state in self.get_device_idle_states(self.measured_cluster):
|
|
if idle_state.index > context.spec.idle_state_index:
|
|
idle_state.disable = 1
|
|
else:
|
|
idle_state.disable = 0
|
|
|
|
def fast_start(self, context): # pylint: disable=unused-argument
|
|
self.start_time = time.time()
|
|
|
|
def fast_stop(self, context): # pylint: disable=unused-argument
|
|
self.run_time = time.time() - self.start_time
|
|
|
|
def on_iteration_start(self, context):
|
|
self.setup_measurement(context.spec.cluster)
|
|
|
|
def thermal_correction(self, context):
|
|
if not self.num_of_freqs_to_thermal_adjust or self.num_of_freqs_to_thermal_adjust > len(self.big_frequencies):
|
|
return 0
|
|
freqs = self.big_frequencies[-self.num_of_freqs_to_thermal_adjust:]
|
|
spec = context.result.spec
|
|
if spec.frequency not in freqs:
|
|
return 0
|
|
data_path = os.path.join(context.output_directory, 'daq', '{}.csv'.format(self.big_core))
|
|
data = pd.read_csv(data_path)['power']
|
|
return _adjust_for_thermal(data, filt_method=lambda x: pd.rolling_median(x, 1000), thresh=0.9, window=5000)
|
|
|
|
# slow to make sure power results have been generated
|
|
def slow_update_result(self, context): # pylint: disable=too-many-branches
|
|
spec = context.result.spec
|
|
cluster = spec.cluster
|
|
is_freq_iteration = spec.label.startswith('freq_')
|
|
perf_metric = 0
|
|
power_metric = 0
|
|
thermal_adjusted_power = 0
|
|
if is_freq_iteration and cluster == 'big':
|
|
thermal_adjusted_power = self.thermal_correction(context)
|
|
for metric in context.result.metrics:
|
|
if metric.name == self.performance_metric:
|
|
perf_metric = metric.value
|
|
elif thermal_adjusted_power and metric.name in self.big_power_metrics:
|
|
power_metric += thermal_adjusted_power * self.power_scaling_factor
|
|
elif (cluster == 'big') and metric.name in self.big_power_metrics:
|
|
power_metric += metric.value * self.power_scaling_factor
|
|
elif (cluster == 'little') and metric.name in self.little_power_metrics:
|
|
power_metric += metric.value * self.power_scaling_factor
|
|
elif thermal_adjusted_power and metric.name in self.big_energy_metrics:
|
|
power_metric += thermal_adjusted_power / self.run_time * self.power_scaling_factor
|
|
elif (cluster == 'big') and metric.name in self.big_energy_metrics:
|
|
power_metric += metric.value / self.run_time * self.power_scaling_factor
|
|
elif (cluster == 'little') and metric.name in self.little_energy_metrics:
|
|
power_metric += metric.value / self.run_time * self.power_scaling_factor
|
|
|
|
if not (power_metric and (perf_metric or not is_freq_iteration)):
|
|
message = 'Incomplete results for {} iteration{}'
|
|
raise InstrumentError(message.format(context.result.spec.id, context.current_iteration))
|
|
|
|
if is_freq_iteration:
|
|
index_matter = [cluster, spec.num_cpus,
|
|
spec.frequency, context.result.iteration]
|
|
data = self.freq_data
|
|
else:
|
|
index_matter = [cluster, spec.num_cpus,
|
|
spec.idle_state_id, spec.idle_state_desc, context.result.iteration]
|
|
data = self.idle_data
|
|
if self.no_hotplug:
|
|
# due to that fact that hotpluging was disabled, power has to be artificially scaled
|
|
# to the number of cores that should have been active if hotplugging had occurred.
|
|
power_metric = spec.num_cpus * (power_metric / self.number_of_cpus[cluster])
|
|
|
|
data.append(index_matter + ['performance', perf_metric])
|
|
data.append(index_matter + ['power', power_metric])
|
|
|
|
def before_overall_results_processing(self, context):
|
|
# pylint: disable=too-many-locals
|
|
if not self.idle_data or not self.freq_data:
|
|
self.logger.warning('Run aborted early; not generating energy_model.')
|
|
return
|
|
output_directory = os.path.join(context.output_directory, 'energy_model')
|
|
os.makedirs(output_directory)
|
|
|
|
df = pd.DataFrame(self.idle_data, columns=['cluster', 'cpus', 'state_id',
|
|
'state', 'iteration', 'metric', 'value'])
|
|
idle_power_table = wa_result_to_power_perf_table(df, '', index=['cluster', 'cpus', 'state'])
|
|
idle_output = os.path.join(output_directory, IDLE_TABLE_FILE)
|
|
with open(idle_output, 'w') as wfh:
|
|
idle_power_table.to_csv(wfh, index=False)
|
|
context.add_artifact('idle_power_table', idle_output, 'export')
|
|
|
|
df = pd.DataFrame(self.freq_data,
|
|
columns=['cluster', 'cpus', 'frequency', 'iteration', 'metric', 'value'])
|
|
freq_power_table = wa_result_to_power_perf_table(df, self.performance_metric,
|
|
index=['cluster', 'cpus', 'frequency'])
|
|
freq_output = os.path.join(output_directory, FREQ_TABLE_FILE)
|
|
with open(freq_output, 'w') as wfh:
|
|
freq_power_table.to_csv(wfh, index=False)
|
|
context.add_artifact('freq_power_table', freq_output, 'export')
|
|
|
|
if self.big_opps is None or self.little_opps is None:
|
|
message = 'OPPs not specified for one or both clusters; cluster power will not be adjusted for leakage.'
|
|
self.logger.warning(message)
|
|
opps = {'big': self.big_opps, 'little': self.little_opps}
|
|
leakages = {'big': self.big_leakage, 'little': self.little_leakage}
|
|
try:
|
|
measured_cpus_table, cpus_table = get_cpus_power_table(freq_power_table, 'frequency', opps, leakages)
|
|
except (ValueError, KeyError, IndexError) as e:
|
|
self.logger.error('Could not create cpu power tables: {}'.format(e))
|
|
return
|
|
measured_cpus_output = os.path.join(output_directory, MEASURED_CPUS_TABLE_FILE)
|
|
with open(measured_cpus_output, 'w') as wfh:
|
|
measured_cpus_table.to_csv(wfh)
|
|
context.add_artifact('measured_cpus_table', measured_cpus_output, 'export')
|
|
cpus_output = os.path.join(output_directory, CPUS_TABLE_FILE)
|
|
with open(cpus_output, 'w') as wfh:
|
|
cpus_table.to_csv(wfh)
|
|
context.add_artifact('cpus_table', cpus_output, 'export')
|
|
|
|
em = build_energy_model(freq_power_table, cpus_table, idle_power_table, self.first_cluster_idle_state)
|
|
em_file = os.path.join(output_directory, '{}_em.c'.format(self.device_name))
|
|
em_text = generate_em_c_file(em, self.big_core, self.little_core,
|
|
self.em_template_file, em_file)
|
|
context.add_artifact('em', em_file, 'data')
|
|
|
|
report_file = os.path.join(output_directory, 'report.html')
|
|
generate_report(freq_power_table, measured_cpus_table, cpus_table,
|
|
idle_power_table, self.report_template_file,
|
|
self.device_name, em_text, report_file)
|
|
context.add_artifact('pm_report', report_file, 'export')
|
|
|
|
def initialize_result_tracking(self):
|
|
self.freq_data = []
|
|
self.idle_data = []
|
|
self.big_power_metrics = []
|
|
self.little_power_metrics = []
|
|
self.big_energy_metrics = []
|
|
self.little_energy_metrics = []
|
|
if self.power_metric:
|
|
self.big_power_metrics = [pm.format(core=self.big_core) for pm in self.power_metric]
|
|
self.little_power_metrics = [pm.format(core=self.little_core) for pm in self.power_metric]
|
|
else: # must be energy_metric
|
|
self.big_energy_metrics = [em.format(core=self.big_core) for em in self.energy_metric]
|
|
self.little_energy_metrics = [em.format(core=self.little_core) for em in self.energy_metric]
|
|
|
|
def configure_clusters(self):
|
|
self.measured_cores = None
|
|
self.measuring_cores = None
|
|
self.cpuset = self.device.get_cgroup_controller('cpuset')
|
|
self.cpuset.create_group('big', self.big_cpus, [0])
|
|
self.cpuset.create_group('little', self.little_cpus, [0])
|
|
for cluster in set(self.device.core_clusters):
|
|
self.device.set_cluster_governor(cluster, 'userspace')
|
|
|
|
def discover_idle_states(self):
|
|
online_cpu = self.device.get_online_cpus(self.big_core)[0]
|
|
self.big_idle_states = self.device.get_cpuidle_states(online_cpu)
|
|
online_cpu = self.device.get_online_cpus(self.little_core)[0]
|
|
self.little_idle_states = self.device.get_cpuidle_states(online_cpu)
|
|
if not (len(self.big_idle_states) >= 2 and len(self.little_idle_states) >= 2):
|
|
raise DeviceError('There do not appeart to be at least two idle states '
|
|
'on at least one of the clusters.')
|
|
|
|
def setup_measurement(self, measured):
|
|
measuring = 'big' if measured == 'little' else 'little'
|
|
self.measured_cluster = measured
|
|
self.measuring_cluster = measuring
|
|
self.measured_cpus = self.big_cpus if measured == 'big' else self.little_cpus
|
|
self.measuring_cpus = self.little_cpus if measured == 'big' else self.big_cpus
|
|
self.reset()
|
|
|
|
def reset(self):
|
|
self.enable_all_cores()
|
|
self.enable_all_idle_states()
|
|
self.reset_cgroups()
|
|
self.cpuset.move_all_tasks_to(self.measuring_cluster)
|
|
server_process = 'adbd' if self.device.platform == 'android' else 'sshd'
|
|
server_pids = self.device.get_pids_of(server_process)
|
|
children_ps = [e for e in self.device.ps()
|
|
if e.ppid in server_pids and e.name != 'sshd']
|
|
children_pids = [e.pid for e in children_ps]
|
|
pids_to_move = server_pids + children_pids
|
|
self.cpuset.root.add_tasks(pids_to_move)
|
|
for pid in pids_to_move:
|
|
try:
|
|
self.device.execute('busybox taskset -p 0x{:x} {}'.format(list_to_mask(self.measuring_cpus), pid))
|
|
except DeviceError:
|
|
pass
|
|
|
|
def enable_all_cores(self):
|
|
counter = Counter(self.device.core_names)
|
|
for core, number in counter.iteritems():
|
|
self.device.set_number_of_online_cpus(core, number)
|
|
self.big_cpus = self.device.get_online_cpus(self.big_core)
|
|
self.little_cpus = self.device.get_online_cpus(self.little_core)
|
|
|
|
def enable_all_idle_states(self):
|
|
for cpu in self.device.online_cpus:
|
|
for state in self.device.get_cpuidle_states(cpu):
|
|
state.disable = 0
|
|
|
|
def reset_cgroups(self):
|
|
self.big_cpus = self.device.get_online_cpus(self.big_core)
|
|
self.little_cpus = self.device.get_online_cpus(self.little_core)
|
|
self.cpuset.big.set(self.big_cpus, 0)
|
|
self.cpuset.little.set(self.little_cpus, 0)
|
|
|
|
def perform_runtime_validation(self):
|
|
if not self.device.is_rooted:
|
|
raise InstrumentError('the device must be rooted to generate energy models')
|
|
if 'userspace' not in self.device.list_available_cluster_governors(0):
|
|
raise InstrumentError('userspace cpufreq governor must be enabled')
|
|
|
|
error_message = 'Frequency {} is not supported by {} cores'
|
|
available_frequencies = self.device.list_available_core_frequencies(self.big_core)
|
|
if self.big_frequencies:
|
|
for freq in self.big_frequencies:
|
|
if freq not in available_frequencies:
|
|
raise ConfigError(error_message.format(freq, self.big_core))
|
|
else:
|
|
self.big_frequencies = available_frequencies
|
|
available_frequencies = self.device.list_available_core_frequencies(self.little_core)
|
|
if self.little_frequencies:
|
|
for freq in self.little_frequencies:
|
|
if freq not in available_frequencies:
|
|
raise ConfigError(error_message.format(freq, self.little_core))
|
|
else:
|
|
self.little_frequencies = available_frequencies
|
|
|
|
def initialize_job_queue(self, context):
|
|
old_specs = []
|
|
for job in context.runner.job_queue:
|
|
if job.spec not in old_specs:
|
|
old_specs.append(job.spec)
|
|
new_specs = self.get_cluster_specs(old_specs, 'big', context)
|
|
new_specs.extend(self.get_cluster_specs(old_specs, 'little', context))
|
|
|
|
# Update config to refect jobs that will actually run.
|
|
context.config.workload_specs = new_specs
|
|
config_file = os.path.join(context.host_working_directory, 'run_config.json')
|
|
with open(config_file, 'wb') as wfh:
|
|
context.config.serialize(wfh)
|
|
|
|
context.runner.init_queue(new_specs)
|
|
|
|
def get_cluster_specs(self, old_specs, cluster, context):
|
|
core = self.get_core_name(cluster)
|
|
self.number_of_cpus[cluster] = sum([1 for c in self.device.core_names if c == core])
|
|
|
|
cluster_frequencies = self.get_frequencies_param(cluster)
|
|
if not cluster_frequencies:
|
|
raise InstrumentError('Could not read available frequencies for {}'.format(core))
|
|
min_frequency = min(cluster_frequencies)
|
|
|
|
idle_states = self.get_device_idle_states(cluster)
|
|
new_specs = []
|
|
for state in idle_states:
|
|
for num_cpus in xrange(1, self.number_of_cpus[cluster] + 1):
|
|
spec = old_specs[0].copy()
|
|
spec.workload_name = self.idle_workload
|
|
spec.workload_parameters = self.idle_workload_params
|
|
spec.idle_state_id = state.id
|
|
spec.idle_state_desc = state.desc
|
|
spec.idle_state_index = state.index
|
|
if not self.no_hotplug:
|
|
spec.runtime_parameters['{}_cores'.format(core)] = num_cpus
|
|
spec.runtime_parameters['{}_frequency'.format(core)] = min_frequency
|
|
spec.runtime_parameters['ui'] = 'off'
|
|
spec.cluster = cluster
|
|
spec.num_cpus = num_cpus
|
|
spec.id = '{}_idle_{}_{}'.format(cluster, state.id, num_cpus)
|
|
spec.label = 'idle_{}'.format(cluster)
|
|
spec.number_of_iterations = old_specs[0].number_of_iterations
|
|
spec.load(self.device, context.config.ext_loader)
|
|
spec.workload.init_resources(context)
|
|
spec.workload.validate()
|
|
new_specs.append(spec)
|
|
for old_spec in old_specs:
|
|
if old_spec.workload_name not in ['sysbench', 'dhrystone']:
|
|
raise ConfigError('Only sysbench and dhrystone workloads currently supported for energy_model generation.')
|
|
for freq in cluster_frequencies:
|
|
for num_cpus in xrange(1, self.number_of_cpus[cluster] + 1):
|
|
spec = old_spec.copy()
|
|
spec.runtime_parameters['{}_frequency'.format(core)] = freq
|
|
if not self.no_hotplug:
|
|
spec.runtime_parameters['{}_cores'.format(core)] = num_cpus
|
|
spec.runtime_parameters['ui'] = 'off'
|
|
spec.id = '{}_{}_{}'.format(cluster, num_cpus, freq)
|
|
spec.label = 'freq_{}_{}'.format(cluster, spec.label)
|
|
spec.workload_parameters['taskset_mask'] = list_to_mask(self.get_cpus(cluster))
|
|
spec.workload_parameters['threads'] = num_cpus
|
|
if old_spec.workload_name == 'sysbench':
|
|
# max_requests set to an arbitrary high values to make sure
|
|
# sysbench runs for full duriation even on highly
|
|
# performant cores.
|
|
spec.workload_parameters['max_requests'] = 10000000
|
|
spec.cluster = cluster
|
|
spec.num_cpus = num_cpus
|
|
spec.frequency = freq
|
|
spec.load(self.device, context.config.ext_loader)
|
|
spec.workload.init_resources(context)
|
|
spec.workload.validate()
|
|
new_specs.append(spec)
|
|
return new_specs
|
|
|
|
def disable_thermal_management(self):
|
|
if self.device.file_exists('/sys/class/thermal/thermal_zone0'):
|
|
tzone_paths = self.device.execute('ls /sys/class/thermal/thermal_zone*')
|
|
for tzpath in tzone_paths.strip().split():
|
|
mode_file = '{}/mode'.format(tzpath)
|
|
if self.device.file_exists(mode_file):
|
|
self.device.set_sysfile_value(mode_file, 'disabled')
|
|
|
|
def get_device_idle_states(self, cluster):
|
|
if cluster == 'big':
|
|
online_cpus = self.device.get_online_cpus(self.big_core)
|
|
else:
|
|
online_cpus = self.device.get_online_cpus(self.little_core)
|
|
idle_states = []
|
|
for cpu in online_cpus:
|
|
idle_states.extend(self.device.get_cpuidle_states(cpu))
|
|
return idle_states
|
|
|
|
def get_core_name(self, cluster):
|
|
if cluster == 'big':
|
|
return self.big_core
|
|
else:
|
|
return self.little_core
|
|
|
|
def get_cpus(self, cluster):
|
|
if cluster == 'big':
|
|
return self.big_cpus
|
|
else:
|
|
return self.little_cpus
|
|
|
|
def get_frequencies_param(self, cluster):
|
|
if cluster == 'big':
|
|
return self.big_frequencies
|
|
else:
|
|
return self.little_frequencies
|
|
|
|
|
|
def _adjust_for_thermal(data, filt_method=lambda x: x, thresh=0.9, window=5000, tdiff_threshold=10000):
|
|
n = filt_method(data)
|
|
n = n[~np.isnan(n)] # pylint: disable=no-member
|
|
|
|
d = np.diff(n) # pylint: disable=no-member
|
|
d = d[~np.isnan(d)] # pylint: disable=no-member
|
|
dmin = min(d)
|
|
dmax = max(d)
|
|
|
|
index_up = np.max((d > dmax * thresh).nonzero()) # pylint: disable=no-member
|
|
index_down = np.min((d < dmin * thresh).nonzero()) # pylint: disable=no-member
|
|
low_average = np.average(n[index_up:index_up + window]) # pylint: disable=no-member
|
|
high_average = np.average(n[index_down - window:index_down]) # pylint: disable=no-member
|
|
if low_average > high_average or index_down - index_up < tdiff_threshold:
|
|
return 0
|
|
else:
|
|
return low_average
|
|
|
|
|
|
if __name__ == '__main__':
|
|
import sys # pylint: disable=wrong-import-position,wrong-import-order
|
|
indir, outdir = sys.argv[1], sys.argv[2]
|
|
device_name = 'odroidxu3'
|
|
big_core = 'a15'
|
|
little_core = 'a7'
|
|
first_cluster_idle_state = -1
|
|
|
|
this_dir = os.path.dirname(__file__)
|
|
report_template_file = os.path.join(this_dir, REPORT_TEMPLATE_FILE)
|
|
em_template_file = os.path.join(this_dir, EM_TEMPLATE_FILE)
|
|
|
|
freq_power_table = pd.read_csv(os.path.join(indir, FREQ_TABLE_FILE))
|
|
measured_cpus_table, cpus_table = pd.read_csv(os.path.join(indir, CPUS_TABLE_FILE), # pylint: disable=unbalanced-tuple-unpacking
|
|
header=range(2), index_col=0)
|
|
idle_power_table = pd.read_csv(os.path.join(indir, IDLE_TABLE_FILE))
|
|
|
|
if not os.path.exists(outdir):
|
|
os.makedirs(outdir)
|
|
report_file = os.path.join(outdir, 'report.html')
|
|
em_file = os.path.join(outdir, '{}_em.c'.format(device_name))
|
|
|
|
em = build_energy_model(freq_power_table, cpus_table,
|
|
idle_power_table, first_cluster_idle_state)
|
|
em_text = generate_em_c_file(em, big_core, little_core,
|
|
em_template_file, em_file)
|
|
generate_report(freq_power_table, measured_cpus_table, cpus_table,
|
|
idle_power_table, report_template_file, device_name,
|
|
em_text, report_file)
|