mirror of
https://github.com/ARM-software/workload-automation.git
synced 2024-10-06 10:51:13 +01:00
c82dd87830
cpuidle module implements cpuidle state discovery, query and manipulation for a Linux device. This replaces the more primitive get_cpuidle_states method of LinuxDevice. Renamed APIs (and added a couple of new ones) to be more consistent: "core" APIs take a core name as the parameter (e.g. "a15") or whatever is listed in core_names for that device. "cluster" APIs take a numeric cluster ID (eg. 0) as the parameter. These get mapped using core_clusters for that device. "cpu" APIs take a cpufreq cpu ID as a parameter. These could be integers, e.g. 0, or full string id, e.g. "cpu0".
378 lines
19 KiB
Python
378 lines
19 KiB
Python
# Copyright 2013-2015 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.
|
|
#
|
|
|
|
import os
|
|
import csv
|
|
import re
|
|
|
|
from wlauto import ResultProcessor, settings, instrumentation
|
|
from wlauto.exceptions import ConfigError, ResultProcessorError
|
|
|
|
|
|
class DVFS(ResultProcessor):
|
|
name = 'dvfs'
|
|
description = """
|
|
Reports DVFS state residency data form ftrace power events.
|
|
|
|
This generates a ``dvfs.csv`` in the top-level results directory that,
|
|
for each workload iteration, reports the percentage of time each CPU core
|
|
spent in each of the DVFS frequency states (P-states), as well as percentage
|
|
of the time spent in idle, during the execution of the workload.
|
|
|
|
.. note:: ``trace-cmd`` instrument *MUST* be enabled in the instrumentation,
|
|
and at least ``'power*'`` events must be enabled.
|
|
|
|
|
|
"""
|
|
|
|
def __init__(self, **kwargs):
|
|
super(DVFS, self).__init__(**kwargs)
|
|
self.device = None
|
|
self.infile = None
|
|
self.outfile = None
|
|
self.current_cluster = None
|
|
self.currentstates_of_clusters = []
|
|
self.current_frequency_of_clusters = []
|
|
self.timestamp = []
|
|
self.state_time_map = {} # hold state at timestamp
|
|
self.cpuid_time_map = {} # hold cpuid at timestamp
|
|
self.cpu_freq_time_spent = {}
|
|
self.cpuids_of_clusters = []
|
|
self.power_state = [0, 1, 2, 3]
|
|
self.UNKNOWNSTATE = 4294967295
|
|
self.multiply_factor = None
|
|
self.corename_of_clusters = []
|
|
self.numberofcores_in_cluster = []
|
|
self.minimum_frequency_cluster = []
|
|
self.idlestate_description = {}
|
|
|
|
def validate(self):
|
|
if not instrumentation.instrument_is_installed('trace-cmd'):
|
|
raise ConfigError('"dvfs" works only if "trace_cmd" in enabled in instrumentation')
|
|
|
|
def initialize(self, context): # pylint: disable=R0912
|
|
self.device = context.device
|
|
if not self.device.has('cpuidle'):
|
|
raise ConfigError('Device does not appear to have cpuidle capability; is the right module installed?')
|
|
if not self.device.core_names:
|
|
message = 'Device does not specify its core types (core_names/core_clusters not set in device_config).'
|
|
raise ResultProcessorError(message)
|
|
number_of_clusters = max(self.device.core_clusters) + 1
|
|
# In IKS devices, actual number of cores is double
|
|
# from what we get from device.number_of_cores
|
|
if self.device.scheduler == 'iks':
|
|
self.multiply_factor = 2
|
|
elif self.device.scheduler == 'unknown':
|
|
# Device doesn't specify its scheduler type. It could be IKS, in
|
|
# which case reporeted values would be wrong, so error out.
|
|
message = ('The Device doesn not specify it\'s scheduler type. If you are '
|
|
'using a generic device interface, please make sure to set the '
|
|
'"scheduler" parameter in the device config.')
|
|
raise ResultProcessorError(message)
|
|
else:
|
|
self.multiply_factor = 1
|
|
# separate out the cores in each cluster
|
|
# It is list of list of cores in cluster
|
|
listof_cores_clusters = []
|
|
for cluster in range(number_of_clusters):
|
|
listof_cores_clusters.append([core for core in self.device.core_clusters if core == cluster])
|
|
# Extract minimum frequency of each cluster and
|
|
# the idle power state with its descriptive name
|
|
#
|
|
total_cores = 0
|
|
current_cores = 0
|
|
for cluster, cores_list in enumerate(listof_cores_clusters):
|
|
self.corename_of_clusters.append(self.device.core_names[total_cores])
|
|
if self.device.scheduler != 'iks':
|
|
self.idlestate_description.update({s.id: s.desc for s in self.device.get_cpuidle_states(total_cores)})
|
|
else:
|
|
self.idlestate_description.update({s.id: s.desc for s in self.device.get_cpuidle_states()})
|
|
total_cores += len(cores_list)
|
|
self.numberofcores_in_cluster.append(len(cores_list))
|
|
for i in range(current_cores, total_cores):
|
|
if i in self.device.active_cpus:
|
|
self.minimum_frequency_cluster.append(int(self.device.get_cpu_min_frequency("cpu{}".format(i))))
|
|
break
|
|
current_cores = total_cores
|
|
length_frequency_cluster = len(self.minimum_frequency_cluster)
|
|
if length_frequency_cluster != number_of_clusters:
|
|
diff = number_of_clusters - length_frequency_cluster
|
|
offline_value = -1
|
|
for i in range(diff):
|
|
if self.device.scheduler != 'iks':
|
|
self.minimum_frequency_cluster.append(offline_value)
|
|
else:
|
|
self.minimum_frequency_cluster.append(self.device.iks_switch_frequency)
|
|
|
|
def process_iteration_result(self, result, context):
|
|
"""
|
|
Parse the trace.txt for each iteration, calculate DVFS residency state/frequencies
|
|
and dump the result in csv and flush the data for next iteration.
|
|
"""
|
|
self.infile = os.path.join(context.output_directory, 'trace.txt')
|
|
if os.path.isfile(self.infile):
|
|
self.logger.debug('Running result_processor "dvfs"')
|
|
self.outfile = os.path.join(settings.output_directory, 'dvfs.csv')
|
|
self.flush_parse_initialize()
|
|
self.calculate()
|
|
self.percentage()
|
|
self.generate_csv(context)
|
|
self.logger.debug('Completed result_processor "dvfs"')
|
|
else:
|
|
self.logger.debug('trace.txt not found.')
|
|
|
|
def flush_parse_initialize(self):
|
|
"""
|
|
Store state, cpu_id for each timestamp from trace.txt and flush all the values for
|
|
next iterations.
|
|
"""
|
|
self.current_cluster = 0
|
|
self.current_frequency_of_clusters = []
|
|
self.timestamp = []
|
|
self.currentstates_of_clusters = []
|
|
self.state_time_map = {}
|
|
self.cpuid_time_map = {}
|
|
self.cpu_freq_time_spent = {}
|
|
self.cpuids_of_clusters = []
|
|
self.parse() # Parse trace.txt generated from trace-cmd instrumentation
|
|
# Initialize the states of each core of clusters and frequency of
|
|
# each clusters with its minimum freq
|
|
# cpu_id is assigned for each of clusters.
|
|
# For IKS devices cpuid remains same in other clusters
|
|
# and for other it will increment by 1
|
|
count = 0
|
|
for cluster, cores_number in enumerate(self.numberofcores_in_cluster):
|
|
self.currentstates_of_clusters.append([-1 for dummy in range(cores_number)])
|
|
self.current_frequency_of_clusters.append(self.minimum_frequency_cluster[cluster])
|
|
if self.device.scheduler == 'iks':
|
|
self.cpuids_of_clusters.append([j for j in range(cores_number)])
|
|
else:
|
|
self.cpuids_of_clusters.append(range(count, count + cores_number))
|
|
count += cores_number
|
|
|
|
# Initialize the time spent in each state/frequency for each core.
|
|
for i in range(self.device.number_of_cores * self.multiply_factor):
|
|
self.cpu_freq_time_spent["cpu{}".format(i)] = {}
|
|
for j in self.unique_freq():
|
|
self.cpu_freq_time_spent["cpu{}".format(i)][j] = 0
|
|
# To determine offline -1 state is added
|
|
offline_value = -1
|
|
self.cpu_freq_time_spent["cpu{}".format(i)][offline_value] = 0
|
|
if 0 not in self.unique_freq():
|
|
self.cpu_freq_time_spent["cpu{}".format(i)][0] = 0
|
|
|
|
def update_cluster_freq(self, state, cpu_id):
|
|
""" Update the cluster frequency and current cluster"""
|
|
# For IKS devices cluster changes only possible when
|
|
# freq changes, for other it is determine by cpu_id.
|
|
if self.device.scheduler != 'iks':
|
|
self.current_cluster = self.get_cluster(cpu_id, state)
|
|
if self.get_state_name(state) == "freqstate":
|
|
self.current_cluster = self.get_cluster(cpu_id, state)
|
|
self.current_frequency_of_clusters[self.current_cluster] = state
|
|
|
|
def get_cluster(self, cpu_id, state):
|
|
# For IKS if current state is greater than switch
|
|
# freq then it is in cluster2 else cluster1
|
|
# For other, Look the current cpu_id and check this id
|
|
# belong to which cluster.
|
|
if self.device.scheduler == 'iks':
|
|
return 1 if state >= self.device.iks_switch_frequency else 0
|
|
else:
|
|
for cluster, cpuids_list in enumerate(self.cpuids_of_clusters):
|
|
if cpu_id in cpuids_list:
|
|
return cluster
|
|
|
|
def get_cluster_freq(self):
|
|
return self.current_frequency_of_clusters[self.current_cluster]
|
|
|
|
def update_state(self, state, cpu_id): # pylint: disable=R0912
|
|
"""
|
|
Update state of each cores in every cluster.
|
|
This is done for each timestamp.
|
|
"""
|
|
POWERDOWN = 2
|
|
offline_value = -1
|
|
# if state is in unknowstate, then change state of current cpu_id
|
|
# with cluster freq of current cluster.
|
|
# if state is in powerstate then change state with that power state.
|
|
if self.get_state_name(state) in ["unknownstate", "powerstate"]:
|
|
for i in range(len(self.cpuids_of_clusters[self.current_cluster])):
|
|
if cpu_id == self.cpuids_of_clusters[self.current_cluster][i]:
|
|
if self.get_state_name(state) == "unknownstate":
|
|
self.currentstates_of_clusters[self.current_cluster][i] = self.current_frequency_of_clusters[self.current_cluster]
|
|
elif self.get_state_name(state) == "powerstate":
|
|
self.currentstates_of_clusters[self.current_cluster][i] = state
|
|
# If state is in freqstate then update the state with current state.
|
|
# For IKS, if all cores is in power down and current state is freqstate
|
|
# then update the all the cores in current cluster to current state
|
|
# and other state cluster changed to Power down.
|
|
if self.get_state_name(state) == "freqstate":
|
|
for i, j in enumerate(self.currentstates_of_clusters[self.current_cluster]):
|
|
if j != offline_value:
|
|
self.currentstates_of_clusters[self.current_cluster][i] = state
|
|
if cpu_id == self.cpuids_of_clusters[self.current_cluster][i]:
|
|
self.currentstates_of_clusters[self.current_cluster][i] = state
|
|
if self.device.scheduler == 'iks':
|
|
check = False # All core in cluster is power down
|
|
for i in range(len(self.currentstates_of_clusters[self.current_cluster])):
|
|
if self.currentstates_of_clusters[self.current_cluster][i] != POWERDOWN:
|
|
check = True
|
|
break
|
|
if not check:
|
|
for i in range(len(self.currentstates_of_clusters[self.current_cluster])):
|
|
self.currentstates_of_clusters[self.current_cluster][i] = self.current_frequency_of_clusters[self.current_cluster]
|
|
for cluster, state_list in enumerate(self.currentstates_of_clusters):
|
|
if cluster != self.current_cluster:
|
|
for j in range(len(state_list)):
|
|
self.currentstates_of_clusters[i][j] = POWERDOWN
|
|
|
|
def unique_freq(self):
|
|
""" Determine the unique Frequency and state"""
|
|
unique_freq = []
|
|
for i in self.timestamp:
|
|
if self.state_time_map[i] not in unique_freq and self.state_time_map[i] != self.UNKNOWNSTATE:
|
|
unique_freq.append(self.state_time_map[i])
|
|
for i in self.minimum_frequency_cluster:
|
|
if i not in unique_freq:
|
|
unique_freq.append(i)
|
|
return unique_freq
|
|
|
|
def parse(self):
|
|
"""
|
|
Parse the trace.txt ::
|
|
|
|
store timestamp, state, cpu_id
|
|
---------------------------------------------------------------------------------
|
|
|timestamp| |state| |cpu_id|
|
|
<idle>-0 [001] 294.554380: cpu_idle: state=4294967295 cpu_id=1
|
|
<idle>-0 [001] 294.554454: power_start: type=1 state=0 cpu_id=1
|
|
<idle>-0 [001] 294.554458: cpu_idle: state=0 cpu_id=1
|
|
<idle>-0 [001] 294.554464: power_end: cpu_id=1
|
|
<idle>-0 [001] 294.554471: cpu_idle: state=4294967295 cpu_id=1
|
|
<idle>-0 [001] 294.554590: power_start: type=1 state=0 cpu_id=1
|
|
<idle>-0 [001] 294.554593: cpu_idle: state=0 cpu_id=1
|
|
<idle>-0 [001] 294.554636: power_end: cpu_id=1
|
|
<idle>-0 [001] 294.554639: cpu_idle: state=4294967295 cpu_id=1
|
|
<idle>-0 [001] 294.554669: power_start: type=1 state=0 cpu_id=1
|
|
|
|
|
|
"""
|
|
pattern = re.compile(r'\s+(?P<time>\S+)\S+\s*(?P<desc>(cpu_idle:|cpu_frequency:))\s*state=(?P<state>\d+)\s*cpu_id=(?P<cpu_id>\d+)')
|
|
start_trace = False
|
|
stop_trace = False
|
|
with open(self.infile, 'r') as f:
|
|
for line in f:
|
|
#Start collecting data from label "TRACE_MARKER_START" and
|
|
#stop with label "TRACE_MARKER_STOP"
|
|
if line.find("TRACE_MARKER_START") != -1:
|
|
start_trace = True
|
|
if line.find("TRACE_MARKER_STOP") != -1:
|
|
stop_trace = True
|
|
if start_trace and not stop_trace:
|
|
match = pattern.search(line)
|
|
if match:
|
|
self.timestamp.append(float(match.group('time')))
|
|
self.state_time_map[float(match.group('time'))] = int(match.group('state'))
|
|
self.cpuid_time_map[float(match.group('time'))] = int(match.group('cpu_id'))
|
|
|
|
def get_state_name(self, state):
|
|
if state in self.power_state:
|
|
return "powerstate"
|
|
elif state == self.UNKNOWNSTATE:
|
|
return "unknownstate"
|
|
else:
|
|
return "freqstate"
|
|
|
|
def populate(self, time1, time2):
|
|
diff = time2 - time1
|
|
for cluster, states_list in enumerate(self.currentstates_of_clusters):
|
|
for k, j in enumerate(states_list):
|
|
if self.device.scheduler == 'iks' and cluster == 1:
|
|
self.cpu_freq_time_spent["cpu{}".format(self.cpuids_of_clusters[cluster][k] + len(self.currentstates_of_clusters[0]))][j] += diff
|
|
else:
|
|
self.cpu_freq_time_spent["cpu{}".format(self.cpuids_of_clusters[cluster][k])][j] += diff
|
|
|
|
def calculate(self):
|
|
for i in range(len(self.timestamp) - 1):
|
|
self.update_cluster_freq(self.state_time_map[self.timestamp[i]], self.cpuid_time_map[self.timestamp[i]])
|
|
self.update_state(self.state_time_map[self.timestamp[i]], self.cpuid_time_map[self.timestamp[i]])
|
|
self.populate(self.timestamp[i], self.timestamp[i + 1])
|
|
|
|
def percentage(self):
|
|
"""Normalize the result with total execution time."""
|
|
temp = self.cpu_freq_time_spent.copy()
|
|
for i in self.cpu_freq_time_spent:
|
|
total = 0
|
|
for j in self.cpu_freq_time_spent[i]:
|
|
total += self.cpu_freq_time_spent[i][j]
|
|
for j in self.cpu_freq_time_spent[i]:
|
|
if total != 0:
|
|
temp[i][j] = self.cpu_freq_time_spent[i][j] * 100 / total
|
|
else:
|
|
temp[i][j] = 0
|
|
return temp
|
|
|
|
def generate_csv(self, context): # pylint: disable=R0912,R0914
|
|
""" generate the '''dvfs.csv''' with the state, frequency and cores """
|
|
temp = self.percentage()
|
|
total_state = self.unique_freq()
|
|
offline_value = -1
|
|
ghz_conversion = 1000000
|
|
mhz_conversion = 1000
|
|
with open(self.outfile, 'a+') as f:
|
|
writer = csv.writer(f, delimiter=',')
|
|
reader = csv.reader(f)
|
|
# Create the header in the format below
|
|
# workload name, iteration, state, A7 CPU0,A7 CPU1,A7 CPU2,A7 CPU3,A15 CPU4,A15 CPU5
|
|
if sum(1 for row in reader) == 0:
|
|
header_row = ['workload', 'iteration', 'state']
|
|
count = 0
|
|
for cluster, states_list in enumerate(self.currentstates_of_clusters):
|
|
for dummy_index in range(len(states_list)):
|
|
header_row.append("{} CPU{}".format(self.corename_of_clusters[cluster], count))
|
|
count += 1
|
|
writer.writerow(header_row)
|
|
if offline_value in total_state:
|
|
total_state.remove(offline_value) # remove the offline state
|
|
for i in sorted(total_state):
|
|
temprow = []
|
|
temprow.extend([context.result.spec.label, context.result.iteration])
|
|
if "state{}".format(i) in self.idlestate_description:
|
|
temprow.append(self.idlestate_description["state{}".format(i)])
|
|
else:
|
|
state_value = float(i)
|
|
if state_value / ghz_conversion >= 1:
|
|
temprow.append("{} Ghz".format(state_value / ghz_conversion))
|
|
else:
|
|
temprow.append("{} Mhz".format(state_value / mhz_conversion))
|
|
for j in range(self.device.number_of_cores * self.multiply_factor):
|
|
temprow.append("{0:.3f}".format(temp["cpu{}".format(j)][i]))
|
|
writer.writerow(temprow)
|
|
check_off = True # Checking whether core is OFFLINE
|
|
for i in range(self.device.number_of_cores * self.multiply_factor):
|
|
temp_val = "{0:.3f}".format(temp["cpu{}".format(i)][offline_value])
|
|
if float(temp_val) > 1:
|
|
check_off = False
|
|
break
|
|
if check_off is False:
|
|
temprow = []
|
|
temprow.extend([context.result.spec.label, context.result.iteration])
|
|
temprow.append("OFFLINE")
|
|
for i in range(self.device.number_of_cores * self.multiply_factor):
|
|
temprow.append("{0:.3f}".format(temp["cpu{}".format(i)][offline_value]))
|
|
writer.writerow(temprow)
|
|
|