def compact_nodelist(node_string):
"""Convert a string of nodes into compact representation.
Wraps ``scontrol show hostlist nid05032,nid05033,...`` to compress a list of
nodes to a Slurm nodelist string. This is effectively the reverse of
``expand_nodelist()``
Args:
node_string (str): Comma-separated list of node names (e.g.,
``nid05032,nid05033,...``)
Returns:
str: The compact representation of `node_string` (e.g.,
``nid0[5032-5159]``)
"""
if not isstr(node_string):
node_string = ','.join(list(node_string))
try:
node_string = subprocess.check_output([SCONTROL, 'show', 'hostlist', node_string]).strip()
except OSError as error:
if error.errno == errno.ENOENT:
# "No such file or directory" from subprocess.check_output
pass
if isstr(node_string):
return node_string
else:
return node_string.decode()
def _recast_keys(self, *target_keys):
"""Convert own keys into native Python objects.
Scan self and convert special keys into native Python objects where
appropriate. If no keys are given, scan everything. Do NOT attempt
to recast anything that is not a string--this is to avoid relying on
expand_nodelist if a key is already recast since expand_nodelist does
not function outside of an environment containing Slurm.
Args:
*target_keys (list, optional): Only convert these keys into native
Python object types. If omitted, convert all keys.
"""
scan_keys = len(target_keys)
for counters in self.values():
# if specific keys were passed, only look for those keys
if scan_keys > 0:
for key in target_keys:
value = counters[key]
if key in _RECAST_KEY_MAP and isstr(value):
counters[key] = _RECAST_KEY_MAP[key][0](value)
# otherwise, attempt to recast every key
else:
for key, value in counters.items():
if key in _RECAST_KEY_MAP and isstr(value):
counters[key] = _RECAST_KEY_MAP[key][0](value)
def _load_subprocess(self, *args):
"""Run a subprocess and pass its stdout to a self-initializing parser
"""
cmd = self.subprocess_cmd
if args:
cmd += args
try:
if self.silent_errors:
with open(os.devnull, 'w') as devnull:
output_str = subprocess.check_output(cmd, stderr=devnull)
else:
output_str = subprocess.check_output(cmd)
except subprocess.CalledProcessError as error:
warnings.warn("%s returned nonzero exit code (%d)" %
(cmd, error.returncode))
output_str = error.output
except OSError as error:
if error.errno == errno.ENOENT:
raise type(error)(error.errno, "%s command not found" %
self.subprocess_cmd[0])
raise
if isstr(output_str):
# Python 2 - subprocess.check_output returns a string
self.load_str(output_str)
else:
# Python 3 - subprocess.check_output returns encoded bytes
self.load_str(output_str.decode())
def expand_nodelist(node_string):
"""Expand Slurm compact nodelist into a set of nodes.
Wraps ``scontrol show hostname nid0[5032-5159]`` to expand a Slurm nodelist
string into a list of nodes.
Args:
node_string (str): Node list in Slurm's compact notation (e.g.,
``nid0[5032-5159]``)
Returns:
set: Set of strings which encode the fully expanded node names contained
in `node_string`.
"""
node_names = set([])
try:
output_str = subprocess.check_output([SCONTROL, 'show', 'hostname', node_string])
except OSError as error:
if error.errno == errno.ENOENT:
raise type(error)(error.errno, "Slurm CLI (%s command) not found" % SCONTROL)
raise
if not isstr(output_str):
output_str = output_str.decode() # for python3
for line in output_str.splitlines():
node_name = line.strip()
if node_name:
node_names.add(node_name)
return node_names
def __repr__(self):
"""Serialize self in a format compatible with ``xtdb2proc``.
Returns the object in the same format as the xtdb2proc output so that
this object can be circularly serialized and deserialized.
Returns:
str: String representation of the processor mapping table in a
format compatible with the output of ``xtdb2proc``.
"""
repr_result = ""
for _, record in self.items():
line = []
for key in self.key_order:
try:
val = record[key]
except KeyError:
sys.stderr.write("key does not appear in all records\n")
raise
# Look at the type of each val and return the correponding string
if val is None:
line.append("%s=null" % key)
elif isstr(val):
line.append("%s='%s'" % (key, val))
else:
line.append("%s=%s" % (key, val))
repr_result += ','.join(line) + "\n"
return repr_result
def archive_ost_data(self, lmtdb):
"""Extract and encode data from LMT's OST_DATA table
Queries the LMT database, interprets resulting rows, and populates a
dictionary of TimeSeries objects with those values.
Args:
lmtdb (LmtDb): database object
"""
dataset_names = [
'datatargets/readbytes', 'datatargets/writebytes',
'fullness/bytes', 'fullness/bytestotal', 'fullness/inodes',
'fullness/inodestotal'
]
self.init_datasets(dataset_names, lmtdb.ost_names)
# Now query the OST_DATA table to get byte counts over the query time range
results, columns = lmtdb.get_ost_data(self.query_start,
self.query_end_plusplus)
# Index the columns to speed up insertion of data
col_map = {}
try:
for db_col in [
'TIMESTAMP', 'OST_ID', 'READ_BYTES', 'WRITE_BYTES',
'KBYTES_USED', 'KBYTES_FREE', 'INODES_USED', 'INODES_FREE'
]:
col_map[db_col] = columns.index(db_col)
except ValueError:
raise ValueError("LMT database schema does not match expectation")
# Loop through all the results of the timeseries query
for row in results:
if isstr(row[col_map['TIMESTAMP']]):
# SQLite stores timestamps as a unicode string
timestamp = datetime.datetime.strptime(
row[col_map['TIMESTAMP']], "%Y-%m-%d %H:%M:%S")
else:
# MySQL timestamps are automatically converted to datetime.datetime
timestamp = row[col_map['TIMESTAMP']]
target_name = lmtdb.ost_id_map[row[col_map['OST_ID']]]
for dataset_name in dataset_names:
target_dbcol = self.config[dataset_name].get('column')
if target_dbcol is not None:
self[dataset_name].insert_element(
timestamp, target_name, row[col_map[target_dbcol]])
elif dataset_name == 'fullness/bytestotal':
self[dataset_name].insert_element(
timestamp, target_name, row[col_map['KBYTES_USED']] +
row[col_map['KBYTES_FREE']])
elif dataset_name == 'fullness/inodestotal':
self[dataset_name].insert_element(
timestamp, target_name, row[col_map['INODES_USED']] +
row[col_map['INODES_FREE']])
else:
errmsg = "%s in self.config but missing 'column' setting" % dataset_name
raise KeyError(errmsg)
def init_config():
"""Loads the global configuration.
Loads the site-wide configuration file, then inspects relevant environment
variables for overrides.
"""
global CONFIG
global PYTOKIO_CONFIG_FILE
global DEFAULT_CONFIG_FILE
# Escape hatch for cases where we want to load the module without initalizing
if os.environ.get('PYTOKIO_SKIP_CONFIG') is not None:
return
# Set the default config path - set here rather than in the global namespace
# so site-specific paths don't get baked into the autodoc documentation
DEFAULT_CONFIG_FILE = os.path.join(
os.path.abspath(os.path.dirname(__file__)), 'site.json')
# Load a pytokio config from a special location
PYTOKIO_CONFIG_FILE = os.environ.get('PYTOKIO_CONFIG', DEFAULT_CONFIG_FILE)
try:
with open(PYTOKIO_CONFIG_FILE, 'rt') as config_file:
if HAVE_YAML:
loaded_config = load_and_expand_yaml(config_file)
else:
loaded_config = json.load(config_file)
except (OSError, IOError):
loaded_config = {}
# Load pytokio config file and convert its keys into a set of constants
for _key, _value in loaded_config.items():
# config keys beginning with an underscore get skipped
if _key.startswith('_'):
pass
# if setting this key will overwrite something already in the tokio.config
# namespace, only overwrite if the existing object is something we probably
# loaded from a json
_old_attribute = CONFIG.get(_key.lower())
if _old_attribute is None \
or isstr(_old_attribute) \
or isinstance(_old_attribute, (dict, list)):
CONFIG[_key.lower()] = _value
# Check for magic environment variables to override the contents of the config
# file at runtime
for _magic_variable in MAGIC_VARIABLES:
_magic_value = os.environ.get("PYTOKIO_" + _magic_variable)
if _magic_value is not None:
try:
_magic_value_decoded = json.loads(_magic_value)
except ValueError:
CONFIG[_magic_variable.lower()] = _magic_value
else:
CONFIG[_magic_variable.lower()] = _magic_value_decoded
def archive_mds_data(self, lmtdb):
"""Extract and encode data from LMT's MDS_DATA table
Queries the LMT database, interprets resulting rows, and populates a
dictionary of TimeSeries objects with those values.
Args:
lmtdb (LmtDb): database object
"""
dataset_names = [
'mdservers/cpuload',
]
self.init_datasets(dataset_names, lmtdb.mds_names)
# Now query the MDS_DATA table to get byte counts over the query time range
results, columns = lmtdb.get_mds_data(self.query_start,
self.query_end_plusplus)
# Index the columns to speed up insertion of data
col_map = {}
try:
for db_col in ['TIMESTAMP', 'MDS_ID', 'PCT_CPU']:
col_map[db_col] = columns.index(db_col)
except ValueError:
raise ValueError("LMT database schema does not match expectation")
# Loop through all the results of the timeseries query
for row in results:
if isstr(row[col_map['TIMESTAMP']]):
# SQLite stores timestamps as a unicode string
timestamp = datetime.datetime.strptime(
row[col_map['TIMESTAMP']], "%Y-%m-%d %H:%M:%S")
else:
# MySQL timestamps are automatically converted to datetime.datetime
timestamp = row[col_map['TIMESTAMP']]
target_name = lmtdb.mds_id_map[row[col_map['MDS_ID']]]
for dataset_name in dataset_names:
target_dbcol = self.config[dataset_name].get('column')
# target_dbcol=PCT_CPU, target_name=snx11025n022
if target_dbcol is not None:
self[dataset_name].insert_element(
timestamp, target_name, row[col_map[target_dbcol]])
else:
errmsg = "%s in self.config but missing 'column' setting" % dataset_name
raise KeyError(errmsg)
def reduce_diff(diff_dict):
"""
Take the raw output of .diff() and aggregate the results of each device
"""
reduced = {
'sum': {},
'min': {},
'max': {},
'count': {},
}
for counters in diff_dict['devices'].values():
for counter, value in counters.items():
if counter not in reduced['count']:
reduced['count'][counter] = 1
new = True
else:
reduced['count'][counter] += 1
new = False
if not isstr(value):
if new:
reduced['sum'][counter] = value
reduced['min'][counter] = value
reduced['max'][counter] = value
else:
reduced['sum'][counter] += value
reduced['min'][counter] = min(value,
reduced['min'][counter])
reduced['max'][counter] = max(value,
reduced['max'][counter])
result = {}
for reduction, counters in reduced.items():
for counter, value in counters.items():
reduced_key = "%s_%s" % (reduction, counter)
result[reduced_key] = value
if reduction == 'sum':
reduced_key = "%s_%s" % ('ave', counter)
result[reduced_key] = float(value) / reduced['count'][counter]
return result
def jobs_running_between(start, end, keys=None):
"""Generate a list of Slurm jobs that ran between a time range
Args:
start (datetime.datetime): Find jobs that ended at or after this time
end (datetime.datetime): Find jobs that started at or before this time
state (str): Any valid sacct state
keys (list): List of Slurm fields to return for each running job
Returns:
tokio.connectors.slurm.Slurm: Slurm object containing jobs whose
runtime overlapped with the start and end times
"""
args = [
'--parsable',
'--starttime', start.strftime("%Y-%m-%dT%H:%M:%S"),
'--endtime', end.strftime("%Y-%m-%dT%H:%M:%S"),
'--state', 'R',
'--allusers',
]
if keys is None:
args += ['--format', ','.join(DEFAULT_KEYS)]
else:
args += ['--format', ','.join(keys)]
try:
print(" ".join([SACCT] + args))
output_str = subprocess.check_output([SACCT] + args)
if not isstr(output_str):
output_str = output_str.decode() # for python3
except OSError as error:
if error.errno == errno.ENOENT:
raise type(error)(error.errno, "Slurm CLI (%s command) not found" % SCONTROL)
raise
return Slurm(from_string=output_str)
def parse_counters_fileobj(fileobj, nodename=None):
"""Convert any output of ISDCT into key-value pairs.
Reads the output of a file-like object which contains the output of a
single isdct command. Understands the output of the following options:
* ``isdct show -smart`` (SMART attributes)
* ``isdct show -sensor`` (device health sensors)
* ``isdct show -performance`` (device performance metrics)
* ``isdct show -a`` (drive info)
Args:
fileobj (file): file-like object containing the output of an ISDCT
command
nodename (str): name of node corresponding to `fileobj`, if known
Returns:
dict: dict of dicts keyed by the device serial number.
"""
data = {}
device_sn = None
parse_mode = 0 # =0 for regular counters, 1 for SMART data
smart_buffer = {}
for line in fileobj.readlines():
if not isstr(line):
line = line.decode().strip()
else:
line = line.strip()
if device_sn is None:
rex_match = re.search(REX_SERIAL_NO, line)
if rex_match is not None:
device_sn = rex_match.group(2)
if nodename is not None:
data['node_name'] = nodename
if rex_match.group(1) == "Intel SSD" \
or rex_match.group(1) == "SMART and Health Information":
parse_mode = 0
elif rex_match.group(1) == "SMART Attributes":
parse_mode = 1
else:
raise Exception("Unknown counter file format")
elif parse_mode == 0 and ':' in line:
key, val = line.split(':', 1)
key = _normalize_key(key)
data[key] = val.strip()
elif parse_mode > 0 and ':' in line:
key, val = line.split(':', 1)
key = _normalize_key(key)
smart_buffer[key] = val.strip()
elif parse_mode > 0 and line.startswith('-') and line.endswith('-'):
for key, val in _rekey_smart_buffer(smart_buffer).items():
key = _normalize_key(key)
data[key] = val
smart_buffer = {'_id': line.split()[1]}
if parse_mode > 0: # flush the last SMART register
for key, val in _rekey_smart_buffer(smart_buffer).items():
key = _normalize_key(key)
data[key] = val
if device_sn is None:
warnings.warn("Could not find device serial number in %s" %
fileobj.name)
return {}
return {device_sn: data}
def _merge_parsed_counters(parsed_counters_list):
"""Merge ISDCT outputs into a single object.
Aggregates counters from each record based on the NVMe device serial number,
with redundant counters being overwritten.
Args:
parsed_counters_list (list): List of parsed ISDCT outputs as dicts.
Each list element is a dict with a single key (a device serial
number) and one or more values; each value is itself a dict of
key-value pairs corresponding to ISDCT/SMART counters from that
device.
Returns:
dict: Dict with keys given by all device serial numbers found in
`parsed_counters_list` and whose values are a dict containing keys
and values representing all unique keys across all elements of
`parsed_counters_list`.
"""
all_data = {}
for parsed_counters in parsed_counters_list:
if not parsed_counters:
continue
elif len(parsed_counters) > 1:
raise Exception(
"Received multiple serial numbers from parse_dct_counters_file"
)
else:
device_sn = next(iter(parsed_counters))
### merge file's counter dict with any previous counters we've parsed
if device_sn not in all_data:
all_data[device_sn] = parsed_counters[device_sn]
else:
all_data[device_sn].update(parsed_counters[device_sn])
### attempt to figure out the type of each counter
new_counters = []
for device_sn, counters in all_data.items():
for counter, value in counters.items():
### first, handle counters that do not have an obvious way to cast
if counter in ("temperature", "throttle_status",
"endurance_analyzer"):
tokens = value.split(None, 1)
if len(tokens) == 2:
new_value, unit = tokens
else:
new_value, unit = tokens[0], None
else:
new_value = value
unit = None
new_value = recast_string(value)
### endurance_analyzer can be numeric or an error string
if counter == 'endurance_analyzer':
if isstr(new_value):
new_value = None
unit = None
elif unit is None:
# Intel reports this counter multiple times using different print formats
unit = 'years'
### Insert the key-value pair and its unit of measurement (if available)
all_data[device_sn][counter] = new_value
if unit is not None:
new_counters.append((device_sn, '%s_unit' % counter, unit))
for (device_sn, counter, value) in new_counters:
all_data[device_sn][counter] = value
return all_data
def diff(self, old_isdct, report_zeros=True):
"""Highlight differences between self and another NerscIsdct.
Subtract each counter for each serial number in this object from its
counterpart in ``old_isdct``. Return the changes in each numeric counter
and any serial numbers that have appeared or disappeared.
Args:
old_isdct (NerscIsdct): object with which we should be compared
report_zeros (bool): If True, report all counters even if they
showed no change. Default is True.
Returns:
dict: Dictionary containing the following keys:
* `added_devices` - device serial numbers which exist in self
but not old_isdct
* `removed_devices` - device serial numbers which do not exist
in self but do in old_isdct
* `devices` - dict keyed by device serial numbers and whose
values are dicts of keys whose values are the difference
between old_isdct and self
"""
result = {
'added_devices': [],
'removed_devices': [],
'devices': {},
}
existing_devices = set([])
for serial_no, counters in self.items():
existing_devices.add(serial_no)
# new devices that are appearing for the first time
if serial_no not in old_isdct:
result['added_devices'].append(serial_no)
continue
# calculate the diff in each counter for this device
diff_dict = {}
for counter, value in counters.items():
if counter not in old_isdct[serial_no]:
warnings.warn("Counter %s does not exist in old_isdct" %
counter)
# just highlight different strings, but ignore
# endurance_analyzer (which can be numeric or string-like)
elif isstr(value) and counter != "endurance_analyzer":
if old_isdct[serial_no][counter] != value:
diff_value = "+++%s ---%s" % (
value, old_isdct[serial_no][counter])
else:
diff_value = ""
if report_zeros or diff_value != "":
diff_dict[counter] = diff_value
# subtract numeric counters
else:
try:
diff_value = value - old_isdct[serial_no][counter]
except TypeError:
### endurance_analyzer can be either numeric (common
### case) or an error string (if the drive is brand
### new); just drop the counter in the non-numeric
### case
if counter == 'endurance_analyzer':
continue
error = "incompatible numeric types for %s/%s: [%s] vs [%s]" % (
serial_no, counter, old_isdct[serial_no][counter],
value)
raise TypeError(error)
if report_zeros or diff_value != 0:
diff_dict[counter] = diff_value
result['devices'][serial_no] = diff_dict
# Look for serial numbers that used to exist but do not appear in self
for serial_no in old_isdct:
if serial_no not in existing_devices:
result['removed_devices'].append(serial_no)
return result
def archive_mds_ops_data(self, lmtdb):
"""Extract and encode data from LMT's MDS_OPS_DATA table
Queries the LMT database, interprets resulting rows, and populates a
dictionary of TimeSeries objects with those values. Avoids JOINing the
MDS_VARIABLE_INFO table and instead uses an internal mapping of
OPERATION_IDs to demultiplex the data in MDS_OPS_DATA into different
HDF5 datasets.
Args:
lmtdb (LmtDb): database object
"""
# mapping between OPERATION_INFO.OPERATION_NAME to HDF5 dataset names
opname_to_dataset_name = {
'open': 'mdtargets/opens',
'close': 'mdtargets/closes',
'mknod': 'mdtargets/mknods',
'link': 'mdtargets/links',
'unlink': 'mdtargets/unlinks',
'mkdir': 'mdtargets/mkdirs',
'rmdir': 'mdtargets/rmdirs',
'rename': 'mdtargets/renames',
'getxattr': 'mdtargets/getxattrs',
'statfs': 'mdtargets/statfss',
'setattr': 'mdtargets/setattrs',
'getattr': 'mdtargets/getattrs',
}
dataset_names = list(opname_to_dataset_name.values())
self.init_datasets(dataset_names, lmtdb.mds_names)
results, columns = lmtdb.get_mds_ops_data(self.query_start,
self.query_end_plusplus)
# Index the columns to speed up insertion of data
col_map = {}
try:
for db_col in ['TIMESTAMP', 'MDS_ID', 'OPERATION_ID', 'SAMPLES']:
col_map[db_col] = columns.index(db_col)
except ValueError:
raise ValueError("LMT database schema does not match expectation")
# Loop through all the results of the timeseries query
for row in results:
if isstr(row[col_map['TIMESTAMP']]):
# SQLite stores timestamps as a unicode string
timestamp = datetime.datetime.strptime(
row[col_map['TIMESTAMP']], "%Y-%m-%d %H:%M:%S")
else:
# MySQL timestamps are automatically converted to datetime.datetime
timestamp = row[col_map['TIMESTAMP']]
# figure out the dataset this row's data will go into (this
# implicitly filters out operations that aren't defined in
# opname_to_dataset_name)
op_name = lmtdb.mds_op_id_map[row[col_map['OPERATION_ID']]]
dataset_name = opname_to_dataset_name.get(op_name)
if dataset_name is None:
continue
# figure out which column (MDS name) this row's data will go into
mds_name = lmtdb.mds_id_map.get(row[col_map['MDS_ID']])
if not mds_name:
errmsg = "unknown MDS_ID %s" % row[col_map['MDS_ID']]
warnings.warn(errmsg)
continue
self[dataset_name].insert_element(timestamp, mds_name,
row[col_map['SAMPLES']])
