def __init__(self, queue_config_mapper, single_mode=False):
AgentBase.__init__(self, single_mode)
self.queueConfigMapper = queue_config_mapper
self.dbProxy = DBProxy()
self.pluginFactory = PluginFactory()
self.startTimestamp = time.time()
self.monitor_fifo = MonitorFIFO()
self.apfmon = Apfmon(self.queueConfigMapper)
def __init__(self, queue_config_mapper, single_mode=False):
AgentBase.__init__(self, single_mode)
self.queueConfigMapper = queue_config_mapper
self.dbProxy = DBProxy()
self.workerMaker = WorkerMaker()
self.workerAdjuster = WorkerAdjuster(queue_config_mapper)
self.pluginFactory = PluginFactory()
self.monitor_fifo = MonitorFIFO()
self.apfmon = Apfmon(self.queueConfigMapper)
def __init__(self, queue_config_mapper):
self.queue_configMapper = queue_config_mapper
self.pluginFactory = PluginFactory()
self.dbProxy = DBProxy()
self.throttlerMap = dict()
self.apf_mon = Apfmon(self.queue_configMapper)
try:
self.maxNewWorkers = harvester_config.submitter.maxNewWorkers
except AttributeError:
self.maxNewWorkers = None
def __init__(self, queue_config_mapper, single_mode=False):
AgentBase.__init__(self, single_mode)
self.queueConfigMapper = queue_config_mapper
self.dbProxy = DBProxy()
self.pluginFactory = PluginFactory()
self.startTimestamp = time.time()
self.monitor_fifo = MonitorFIFO()
if self.monitor_fifo.enabled:
self.monitor_event_fifo = MonitorEventFIFO()
else:
self.monitor_event_fifo = None
self.apfmon = Apfmon(self.queueConfigMapper)
self.eventBasedMonCoreList = []
if getattr(harvester_config.monitor, 'eventBasedEnable', False):
for pluginConf in harvester_config.monitor.eventBasedPlugins:
pluginFactory = PluginFactory()
self.eventBasedMonCoreList.append(pluginFactory.get_plugin(pluginConf))
def __init__(self, queue_config_mapper):
self.queueConfigMapper = queue_config_mapper
self.pluginFactory = PluginFactory()
self.dbProxy = DBProxy()
self.throttlerMap = dict()
self.apf_mon = Apfmon(self.queueConfigMapper)
try:
self.maxNewWorkers = harvester_config.submitter.maxNewWorkers
except AttributeError:
self.maxNewWorkers = None
class WorkerAdjuster(object):
# constructor
def __init__(self, queue_config_mapper):
self.queue_configMapper = queue_config_mapper
self.pluginFactory = PluginFactory()
self.dbProxy = DBProxy()
self.throttlerMap = dict()
self.apf_mon = Apfmon(self.queue_configMapper)
try:
self.maxNewWorkers = harvester_config.submitter.maxNewWorkers
except AttributeError:
self.maxNewWorkers = None
# define number of workers to submit based on various information
def define_num_workers(self, static_num_workers, site_name):
tmp_log = core_utils.make_logger(_logger,
'site={0}'.format(site_name),
method_name='define_num_workers')
tmp_log.debug('start')
tmp_log.debug('static_num_workers: {0}'.format(static_num_workers))
dyn_num_workers = copy.deepcopy(static_num_workers)
try:
# get queue status
queue_stat = self.dbProxy.get_cache("panda_queues.json", None)
if queue_stat is None:
queue_stat = dict()
else:
queue_stat = queue_stat.data
# get job statistics
job_stats = self.dbProxy.get_cache("job_statistics.json", None)
if job_stats is None:
job_stats = dict()
else:
job_stats = job_stats.data
# define num of new workers
for queue_name in static_num_workers:
# get queue
queue_config = self.queue_configMapper.get_queue(queue_name)
worker_limits_dict = self.dbProxy.get_worker_limits(queue_name)
max_workers = worker_limits_dict.get('maxWorkers', 0)
n_queue_limit = worker_limits_dict.get('nQueueLimitWorker', 0)
n_queue_limit_per_rt = worker_limits_dict[
'nQueueLimitWorkerPerRT']
n_queue_total, n_ready_total, n_running_total = 0, 0, 0
apf_msg = None
apf_data = None
for job_type, jt_values in iteritems(
static_num_workers[queue_name]):
for resource_type, tmp_val in iteritems(jt_values):
tmp_log.debug(
'Processing queue {0} job_type {1} resource_type {2} with static_num_workers {3}'
.format(queue_name, job_type, resource_type,
tmp_val))
# set 0 to num of new workers when the queue is disabled
if queue_name in queue_stat and queue_stat[queue_name][
'status'] in [
'offline', 'standby', 'maintenance'
]:
dyn_num_workers[queue_name][job_type][
resource_type]['nNewWorkers'] = 0
ret_msg = 'set n_new_workers=0 since status={0}'.format(
queue_stat[queue_name]['status'])
tmp_log.debug(ret_msg)
apf_msg = 'Not submitting workers since queue status = {0}'.format(
queue_stat[queue_name]['status'])
continue
# protection against not-up-to-date queue config
if queue_config is None:
dyn_num_workers[queue_name][job_type][
resource_type]['nNewWorkers'] = 0
ret_msg = 'set n_new_workers=0 due to missing queue_config'
tmp_log.debug(ret_msg)
apf_msg = 'Not submitting workers because of missing queue_config'
continue
# get throttler
if queue_name not in self.throttlerMap:
if hasattr(queue_config, 'throttler'):
throttler = self.pluginFactory.get_plugin(
queue_config.throttler)
else:
throttler = None
self.throttlerMap[queue_name] = throttler
# check throttler
throttler = self.throttlerMap[queue_name]
if throttler is not None:
to_throttle, tmp_msg = throttler.to_be_throttled(
queue_config)
if to_throttle:
dyn_num_workers[queue_name][job_type][
resource_type]['nNewWorkers'] = 0
ret_msg = 'set n_new_workers=0 by {0}:{1}'.format(
throttler.__class__.__name__, tmp_msg)
tmp_log.debug(ret_msg)
continue
# check stats
n_queue = tmp_val['nQueue']
n_ready = tmp_val['nReady']
n_running = tmp_val['nRunning']
if resource_type != 'ANY' and job_type != 'ANY' and job_type is not None:
n_queue_total += n_queue
n_ready_total += n_ready
n_running_total += n_running
if queue_config.runMode == 'slave':
n_new_workers_def = tmp_val['nNewWorkers']
if n_new_workers_def == 0:
dyn_num_workers[queue_name][job_type][
resource_type]['nNewWorkers'] = 0
ret_msg = 'set n_new_workers=0 by panda in slave mode'
tmp_log.debug(ret_msg)
continue
else:
n_new_workers_def = None
# define num of new workers based on static site config
n_new_workers = 0
if n_queue >= n_queue_limit_per_rt > 0:
# enough queued workers
ret_msg = 'No n_new_workers since n_queue({0})>=n_queue_limit_per_rt({1})'.format(
n_queue, n_queue_limit_per_rt)
tmp_log.debug(ret_msg)
pass
elif (n_queue + n_ready +
n_running) >= max_workers > 0:
# enough workers in the system
ret_msg = 'No n_new_workers since n_queue({0}) + n_ready({1}) + n_running({2}) '.format(
n_queue, n_ready, n_running)
ret_msg += '>= max_workers({0})'.format(
max_workers)
tmp_log.debug(ret_msg)
pass
else:
max_queued_workers = None
if n_queue_limit_per_rt > 0: # there is a limit set for the queue
max_queued_workers = n_queue_limit_per_rt
# Reset the maxQueueWorkers according to particular
if n_new_workers_def is not None: # don't surpass limits given centrally
maxQueuedWorkers_slave = n_new_workers_def + n_queue
if max_queued_workers is not None:
max_queued_workers = min(
maxQueuedWorkers_slave,
max_queued_workers)
else:
max_queued_workers = maxQueuedWorkers_slave
elif queue_config.mapType == 'NoJob': # for pull mode, limit to activated jobs
# limit the queue to the number of activated jobs to avoid empty pilots
try:
n_activated = max(
job_stats[queue_name]['activated'],
1) # avoid no activity queues
queue_limit = max_queued_workers
max_queued_workers = min(
n_activated, max_queued_workers)
tmp_log.debug(
'limiting max_queued_workers to min(n_activated={0}, queue_limit={1})'
.format(n_activated, queue_limit))
except KeyError:
tmp_log.warning(
'n_activated not defined, defaulting to configured queue limits'
)
pass
if max_queued_workers is None: # no value found, use default value
max_queued_workers = 1
# new workers
n_new_workers = max(max_queued_workers - n_queue,
0)
tmp_log.debug(
'setting n_new_workers to {0} in max_queued_workers calculation'
.format(n_new_workers))
if max_workers > 0:
n_new_workers = min(
n_new_workers,
max(
max_workers - n_queue - n_ready -
n_running, 0))
tmp_log.debug(
'setting n_new_workers to {0} to respect max_workers'
.format(n_new_workers))
if queue_config.maxNewWorkersPerCycle > 0:
n_new_workers = min(
n_new_workers,
queue_config.maxNewWorkersPerCycle)
tmp_log.debug(
'setting n_new_workers to {0} in order to respect maxNewWorkersPerCycle'
.format(n_new_workers))
if self.maxNewWorkers is not None and self.maxNewWorkers > 0:
n_new_workers = min(n_new_workers,
self.maxNewWorkers)
tmp_log.debug(
'setting n_new_workers to {0} in order to respect universal maxNewWorkers'
.format(n_new_workers))
dyn_num_workers[queue_name][job_type][resource_type][
'nNewWorkers'] = n_new_workers
# adjust n_new_workers for UCORE to let aggregations over RT respect nQueueLimitWorker and max_workers
if queue_config is None:
max_new_workers_per_cycle = 0
ret_msg = 'set max_new_workers_per_cycle=0 in UCORE aggregation due to missing queue_config'
tmp_log.debug(ret_msg)
else:
max_new_workers_per_cycle = queue_config.maxNewWorkersPerCycle
if len(dyn_num_workers[queue_name]) > 1:
total_new_workers_rts = 0
for _jt in dyn_num_workers[queue_name]:
for _rt in dyn_num_workers[queue_name][_jt]:
if _jt != 'ANY' and _rt != 'ANY':
total_new_workers_rts = total_new_workers_rts + dyn_num_workers[
queue_name][_jt][_rt]['nNewWorkers']
n_new_workers_max_agg = min(
max(n_queue_limit - n_queue_total, 0),
max(
max_workers - n_queue_total - n_ready_total -
n_running_total, 0))
if max_new_workers_per_cycle >= 0:
n_new_workers_max_agg = min(n_new_workers_max_agg,
max_new_workers_per_cycle)
if self.maxNewWorkers is not None and self.maxNewWorkers > 0:
n_new_workers_max_agg = min(n_new_workers_max_agg,
self.maxNewWorkers)
# exceeded max, to adjust
if total_new_workers_rts > n_new_workers_max_agg:
if n_new_workers_max_agg == 0:
for job_type in dyn_num_workers[queue_name]:
for resource_type in dyn_num_workers[
queue_name][job_type]:
dyn_num_workers[queue_name][job_type][
resource_type]['nNewWorkers'] = 0
tmp_log.debug(
'No n_new_workers since n_new_workers_max_agg=0 for UCORE'
)
else:
tmp_log.debug(
'n_new_workers_max_agg={0} for UCORE'.format(
n_new_workers_max_agg))
_d = dyn_num_workers[queue_name].copy()
del _d['ANY']
# TODO: needs to be recalculated
simple_rt_nw_list = []
for job_type in _d: # jt: job type
for resource_type in _d[
job_type]: # rt: resource type
simple_rt_nw_list.append([
(resource_type, job_type),
_d[job_type][resource_type].get(
'nNewWorkers', 0), 0
])
_countdown = n_new_workers_max_agg
for _rt_list in simple_rt_nw_list:
(resource_type,
job_type), n_new_workers_orig, _r = _rt_list
n_new_workers, remainder = divmod(
n_new_workers_orig * n_new_workers_max_agg,
total_new_workers_rts)
dyn_num_workers[queue_name][
job_type].setdefault(
resource_type, {
'nReady': 0,
'nRunning': 0,
'nQueue': 0,
'nNewWorkers': 0
})
dyn_num_workers[queue_name][job_type][
resource_type][
'nNewWorkers'] = n_new_workers
_rt_list[2] = remainder
_countdown -= n_new_workers
_s_list = sorted(simple_rt_nw_list,
key=(lambda x: x[1]))
sorted_rt_nw_list = sorted(_s_list,
key=(lambda x: x[2]),
reverse=True)
for (
resource_type, job_type
), n_new_workers_orig, remainder in sorted_rt_nw_list:
if _countdown <= 0:
break
dyn_num_workers[queue_name][job_type][
resource_type]['nNewWorkers'] += 1
_countdown -= 1
for job_type in dyn_num_workers[queue_name]:
for resource_type in dyn_num_workers[queue_name][
job_type]:
if job_type == 'ANY' or resource_type == 'ANY':
continue
n_new_workers = dyn_num_workers[queue_name][
job_type][resource_type]['nNewWorkers']
tmp_log.debug(
'setting n_new_workers to {0} of job_type {1} resource_type {2} in order to respect RT aggregations for UCORE'
.format(n_new_workers, job_type,
resource_type))
if not apf_msg:
apf_data = copy.deepcopy(dyn_num_workers[queue_name])
self.apf_mon.update_label(queue_name, apf_msg, apf_data)
# dump
tmp_log.debug('defined {0}'.format(str(dyn_num_workers)))
return dyn_num_workers
except Exception:
# dump error
err_msg = core_utils.dump_error_message(tmp_log)
return None
class Monitor(AgentBase):
# constructor
def __init__(self, queue_config_mapper, single_mode=False):
AgentBase.__init__(self, single_mode)
self.queueConfigMapper = queue_config_mapper
self.dbProxy = DBProxy()
self.pluginFactory = PluginFactory()
self.startTimestamp = time.time()
self.monitor_fifo = MonitorFIFO()
self.apfmon = Apfmon(self.queueConfigMapper)
# main loop
def run(self):
lockedBy = 'monitor-{0}'.format(self.get_pid())
# init messengers
for queueConfig in self.queueConfigMapper.get_all_queues().values():
# just import for module initialization
self.pluginFactory.get_plugin(queueConfig.messenger)
# main
try:
fifoSleepTimeMilli = harvester_config.monitor.fifoSleepTimeMilli
except AttributeError:
fifoSleepTimeMilli = 5000
try:
fifoCheckDuration = harvester_config.monitor.fifoCheckDuration
except AttributeError:
fifoCheckDuration = 30
try:
fifoMaxWorkersPerChunk = harvester_config.monitor.fifoMaxWorkersPerChunk
except AttributeError:
fifoMaxWorkersPerChunk = 500
try:
fifoProtectiveDequeue = harvester_config.monitor.fifoProtectiveDequeue
except AttributeError:
fifoProtectiveDequeue = True
last_DB_cycle_timestamp = 0
monitor_fifo = self.monitor_fifo
sleepTime = (fifoSleepTimeMilli / 1000.0) \
if monitor_fifo.enabled else harvester_config.monitor.sleepTime
adjusted_sleepTime = sleepTime
if monitor_fifo.enabled:
monitor_fifo.restore()
while True:
sw_main = core_utils.get_stopwatch()
mainLog = self.make_logger(_logger,
'id={0}'.format(lockedBy),
method_name='run')
mainLog.debug('start a monitor cycle')
if time.time() >= last_DB_cycle_timestamp + harvester_config.monitor.sleepTime and \
not (monitor_fifo.enabled and self.singleMode):
# run with workers from DB
sw_db = core_utils.get_stopwatch()
mainLog.debug('starting run with DB')
mainLog.debug('getting workers to monitor')
workSpecsPerQueue = self.dbProxy.get_workers_to_update(
harvester_config.monitor.maxWorkers,
harvester_config.monitor.checkInterval,
harvester_config.monitor.lockInterval, lockedBy)
mainLog.debug('got {0} queues'.format(len(workSpecsPerQueue)))
# loop over all workers
for queueName, configIdWorkSpecs in iteritems(
workSpecsPerQueue):
for configID, workSpecsList in iteritems(
configIdWorkSpecs):
retVal = self.monitor_agent_core(lockedBy,
queueName,
workSpecsList,
config_id=configID)
if monitor_fifo.enabled and retVal is not None:
workSpecsToEnqueue, workSpecsToEnqueueToHead, timeNow_timestamp, fifoCheckInterval = retVal
if workSpecsToEnqueue:
mainLog.debug('putting workers to FIFO')
try:
score = fifoCheckInterval + timeNow_timestamp
monitor_fifo.put(
(queueName, workSpecsToEnqueue), score)
mainLog.info(
'put workers of {0} to FIFO with score {1}'
.format(queueName, score))
except Exception as errStr:
mainLog.error(
'failed to put object from FIFO: {0}'.
format(errStr))
if workSpecsToEnqueueToHead:
mainLog.debug('putting workers to FIFO head')
try:
score = fifoCheckInterval - timeNow_timestamp
monitor_fifo.put(
(queueName, workSpecsToEnqueueToHead),
score)
mainLog.info(
'put workers of {0} to FIFO with score {1}'
.format(queueName, score))
except Exception as errStr:
mainLog.error(
'failed to put object from FIFO head: {0}'
.format(errStr))
last_DB_cycle_timestamp = time.time()
if sw_db.get_elapsed_time_in_sec(
) > harvester_config.monitor.lockInterval:
mainLog.warning(
'a single DB cycle was longer than lockInterval ' +
sw_db.get_elapsed_time())
else:
mainLog.debug('done a DB cycle' + sw_db.get_elapsed_time())
mainLog.debug('ended run with DB')
elif monitor_fifo.enabled:
# run with workers from FIFO
sw = core_utils.get_stopwatch()
n_loops = 0
n_loops_hit = 0
last_fifo_cycle_timestamp = time.time()
to_break = False
obj_dequeued_id_list = []
obj_to_enqueue_dict = collections.defaultdict(
lambda: [[], 0, 0])
obj_to_enqueue_to_head_dict = collections.defaultdict(
lambda: [[], 0, 0])
remaining_obj_to_enqueue_dict = {}
remaining_obj_to_enqueue_to_head_dict = {}
n_chunk_peeked_stat, sum_overhead_time_stat = 0, 0.0
while time.time(
) < last_fifo_cycle_timestamp + fifoCheckDuration:
sw.reset()
n_loops += 1
retVal, overhead_time = monitor_fifo.to_check_workers()
if overhead_time is not None:
n_chunk_peeked_stat += 1
sum_overhead_time_stat += overhead_time
if retVal:
# check fifo size
fifo_size = monitor_fifo.size()
mainLog.debug('FIFO size is {0}'.format(fifo_size))
mainLog.debug('starting run with FIFO')
try:
obj_gotten = monitor_fifo.get(
timeout=1, protective=fifoProtectiveDequeue)
except Exception as errStr:
mainLog.error(
'failed to get object from FIFO: {0}'.format(
errStr))
else:
if obj_gotten is not None:
sw_fifo = core_utils.get_stopwatch()
if fifoProtectiveDequeue:
obj_dequeued_id_list.append(obj_gotten.id)
queueName, workSpecsList = obj_gotten.item
mainLog.debug(
'got a chunk of {0} workers of {1} from FIFO'
.format(len(workSpecsList), queueName) +
sw.get_elapsed_time())
sw.reset()
configID = None
for workSpecs in workSpecsList:
if configID is None and len(workSpecs) > 0:
configID = workSpecs[0].configID
for workSpec in workSpecs:
if workSpec.pandaid_list is None:
_jobspec_list = workSpec.get_jobspec_list(
)
if _jobspec_list is not None:
workSpec.pandaid_list = [
j.PandaID
for j in workSpec.
get_jobspec_list()
]
else:
workSpec.pandaid_list = []
workSpec.force_update(
'pandaid_list')
retVal = self.monitor_agent_core(
lockedBy,
queueName,
workSpecsList,
from_fifo=True,
config_id=configID)
if retVal is not None:
workSpecsToEnqueue, workSpecsToEnqueueToHead, timeNow_timestamp, fifoCheckInterval = retVal
try:
if len(obj_to_enqueue_dict[queueName]
[0]) + len(
workSpecsToEnqueue
) <= fifoMaxWorkersPerChunk:
obj_to_enqueue_dict[queueName][
0].extend(workSpecsToEnqueue)
obj_to_enqueue_dict[queueName][
1] = max(
obj_to_enqueue_dict[
queueName][1],
timeNow_timestamp)
obj_to_enqueue_dict[queueName][
2] = max(
obj_to_enqueue_dict[
queueName][2],
fifoCheckInterval)
else:
to_break = True
remaining_obj_to_enqueue_dict[
queueName] = [
workSpecsToEnqueue,
timeNow_timestamp,
fifoCheckInterval
]
except Exception as errStr:
mainLog.error(
'failed to gather workers for FIFO: {0}'
.format(errStr))
to_break = True
try:
if len(obj_to_enqueue_to_head_dict[
queueName][0]) + len(
workSpecsToEnqueueToHead
) <= fifoMaxWorkersPerChunk:
obj_to_enqueue_to_head_dict[
queueName][0].extend(
workSpecsToEnqueueToHead)
obj_to_enqueue_to_head_dict[
queueName][1] = max(
obj_to_enqueue_to_head_dict[
queueName][1],
timeNow_timestamp)
obj_to_enqueue_to_head_dict[
queueName][2] = max(
obj_to_enqueue_to_head_dict[
queueName][2],
fifoCheckInterval)
else:
to_break = True
remaining_obj_to_enqueue_to_head_dict[
queueName] = [
workSpecsToEnqueueToHead,
timeNow_timestamp,
fifoCheckInterval
]
except Exception as errStr:
mainLog.error(
'failed to gather workers for FIFO head: {0}'
.format(errStr))
to_break = True
mainLog.debug(
'checked {0} workers from FIFO'.format(
len(workSpecsList)) +
sw.get_elapsed_time())
else:
mainLog.debug(
'monitor_agent_core returned None. Skipped putting to FIFO'
)
if sw_fifo.get_elapsed_time_in_sec(
) > harvester_config.monitor.lockInterval:
mainLog.warning(
'a single FIFO cycle was longer than lockInterval '
+ sw_fifo.get_elapsed_time())
else:
mainLog.debug('done a FIFO cycle' +
sw_fifo.get_elapsed_time())
n_loops_hit += 1
if to_break:
break
else:
mainLog.debug('got nothing in FIFO')
else:
mainLog.debug(
'workers in FIFO too young to check. Skipped')
if self.singleMode:
break
if overhead_time is not None:
time.sleep(
max(-overhead_time * random.uniform(0.1, 1),
adjusted_sleepTime))
else:
time.sleep(
max(fifoCheckDuration * random.uniform(0.1, 1),
adjusted_sleepTime))
mainLog.debug(
'run {0} loops, including {1} FIFO cycles'.format(
n_loops, n_loops_hit))
# enqueue to fifo
sw.reset()
n_chunk_put = 0
mainLog.debug('putting worker chunks to FIFO')
for _dct in (obj_to_enqueue_dict,
remaining_obj_to_enqueue_dict):
for queueName, obj_to_enqueue in iteritems(_dct):
try:
workSpecsToEnqueue, timeNow_timestamp, fifoCheckInterval = obj_to_enqueue
if workSpecsToEnqueue:
score = fifoCheckInterval + timeNow_timestamp
monitor_fifo.put(
(queueName, workSpecsToEnqueue), score)
n_chunk_put += 1
mainLog.info(
'put a chunk of {0} workers of {1} to FIFO with score {2}'
.format(len(workSpecsToEnqueue), queueName,
score))
except Exception as errStr:
mainLog.error(
'failed to put object from FIFO: {0}'.format(
errStr))
mainLog.debug('putting worker chunks to FIFO head')
for _dct in (obj_to_enqueue_to_head_dict,
remaining_obj_to_enqueue_to_head_dict):
for queueName, obj_to_enqueue_to_head in iteritems(_dct):
try:
workSpecsToEnqueueToHead, timeNow_timestamp, fifoCheckInterval = obj_to_enqueue_to_head
if workSpecsToEnqueueToHead:
score = fifoCheckInterval + timeNow_timestamp - 2**32
monitor_fifo.put(
(queueName, workSpecsToEnqueueToHead),
score)
n_chunk_put += 1
mainLog.info(
'put a chunk of {0} workers of {1} to FIFO with score {2}'
.format(len(workSpecsToEnqueueToHead),
queueName, score))
except Exception as errStr:
mainLog.error(
'failed to put object from FIFO head: {0}'.
format(errStr))
# release protective dequeued objects
if fifoProtectiveDequeue and len(obj_dequeued_id_list) > 0:
monitor_fifo.release(ids=obj_dequeued_id_list)
mainLog.debug(
'put {0} worker chunks into FIFO'.format(n_chunk_put) +
sw.get_elapsed_time())
# adjust adjusted_sleepTime
if n_chunk_peeked_stat > 0 and sum_overhead_time_stat > sleepTime:
speedup_factor = (sum_overhead_time_stat - sleepTime) / (
n_chunk_peeked_stat *
harvester_config.monitor.checkInterval)
speedup_factor = max(speedup_factor, 0)
adjusted_sleepTime = adjusted_sleepTime / (1. +
speedup_factor)
elif n_chunk_peeked_stat == 0 or sum_overhead_time_stat < 0:
adjusted_sleepTime = (sleepTime + adjusted_sleepTime) / 2
mainLog.debug('adjusted_sleepTime becomes {0:.3f} sec'.format(
adjusted_sleepTime))
# end run with fifo
mainLog.debug('ended run with FIFO')
# time the cycle
mainLog.debug('done a monitor cycle' + sw_main.get_elapsed_time())
# check if being terminated
if self.terminated(adjusted_sleepTime):
mainLog.debug('terminated')
return
# core of monitor agent to check workers in workSpecsList of queueName
def monitor_agent_core(self,
lockedBy,
queueName,
workSpecsList,
from_fifo=False,
config_id=None):
tmpQueLog = self.make_logger(_logger,
'id={0} queue={1}'.format(
lockedBy, queueName),
method_name='run')
# check queue
if not self.queueConfigMapper.has_queue(queueName, config_id):
tmpQueLog.error('config not found')
return None
# get queue
queueConfig = self.queueConfigMapper.get_queue(queueName, config_id)
try:
apfmon_status_updates = self.queueConfigMapper.queueConfig[
queueName].monitor['apfmon_status_updates']
except Exception:
apfmon_status_updates = False
tmpQueLog.debug(
'apfmon_status_updates: {0}'.format(apfmon_status_updates))
# get plugins
monCore = self.pluginFactory.get_plugin(queueConfig.monitor)
messenger = self.pluginFactory.get_plugin(queueConfig.messenger)
# workspec chunk of active workers
workSpecsToEnqueue_dict = {}
workSpecsToEnqueueToHead_dict = {}
timeNow_timestamp = time.time()
# get fifoCheckInterval for PQ and other fifo attributes
try:
fifoCheckInterval = monCore.fifoCheckInterval
except Exception:
if hasattr(harvester_config.monitor, 'fifoCheckInterval'):
fifoCheckInterval = harvester_config.monitor.fifoCheckInterval
else:
fifoCheckInterval = harvester_config.monitor.checkInterval
try:
forceEnqueueInterval = harvester_config.monitor.fifoForceEnqueueInterval
except AttributeError:
forceEnqueueInterval = 3600
try:
fifoMaxPreemptInterval = harvester_config.monitor.fifoMaxPreemptInterval
except AttributeError:
fifoMaxPreemptInterval = 60
# check workers
allWorkers = [item for sublist in workSpecsList for item in sublist]
tmpQueLog.debug('checking {0} workers'.format(len(allWorkers)))
tmpStat, tmpRetMap = self.check_workers(monCore, messenger, allWorkers,
queueConfig, tmpQueLog,
from_fifo)
if tmpStat:
# loop over all worker chunks
tmpQueLog.debug('update jobs and workers')
iWorker = 0
for workSpecs in workSpecsList:
jobSpecs = None
pandaIDsList = []
eventsToUpdateList = []
filesToStageOutList = dict()
isCheckedList = []
mapType = workSpecs[0].mapType
# loop over workSpecs
for workSpec in workSpecs:
tmpLog = self.make_logger(_logger,
'id={0} workerID={1}'.format(
lockedBy, workSpec.workerID),
method_name='run')
tmpOut = tmpRetMap[workSpec.workerID]
oldStatus = tmpOut['oldStatus']
newStatus = tmpOut['newStatus']
monStatus = tmpOut['monStatus']
diagMessage = tmpOut['diagMessage']
workAttributes = tmpOut['workAttributes']
eventsToUpdate = tmpOut['eventsToUpdate']
filesToStageOut = tmpOut['filesToStageOut']
eventsRequestParams = tmpOut['eventsRequestParams']
nJobsToReFill = tmpOut['nJobsToReFill']
pandaIDs = tmpOut['pandaIDs']
isChecked = tmpOut['isChecked']
tmpStr = 'newStatus={0} monitoredStatus={1} diag={2} '
tmpStr += 'postProcessed={3} files={4}'
tmpLog.debug(
tmpStr.format(newStatus, monStatus, diagMessage,
workSpec.is_post_processed(),
str(filesToStageOut)))
iWorker += 1
# check status
if newStatus not in WorkSpec.ST_LIST:
tmpLog.error('unknown status={0}'.format(newStatus))
return
# update worker
workSpec.set_status(newStatus)
workSpec.set_work_attributes(workAttributes)
workSpec.set_dialog_message(diagMessage)
if isChecked:
workSpec.checkTime = datetime.datetime.utcnow()
isCheckedList.append(isChecked)
if monStatus == WorkSpec.ST_failed:
if not workSpec.has_pilot_error():
workSpec.set_pilot_error(
PilotErrors.ERR_GENERALERROR, diagMessage)
elif monStatus == WorkSpec.ST_cancelled:
if not workSpec.has_pilot_error():
workSpec.set_pilot_error(PilotErrors.ERR_PANDAKILL,
diagMessage)
if monStatus in [
WorkSpec.ST_finished, WorkSpec.ST_failed,
WorkSpec.ST_cancelled
]:
workSpec.set_work_params({'finalMonStatus': monStatus})
# request events
if eventsRequestParams != {}:
workSpec.eventsRequest = WorkSpec.EV_requestEvents
workSpec.eventsRequestParams = eventsRequestParams
# jobs to refill
if nJobsToReFill is not None:
workSpec.nJobsToReFill = nJobsToReFill
# get associated jobs for the worker chunk
if workSpec.hasJob == 1 and jobSpecs is None:
jobSpecs = self.dbProxy.get_jobs_with_worker_id(
workSpec.workerID,
None,
only_running=True,
slim=True)
# pandaIDs for push
pandaIDsList.append(pandaIDs)
if len(eventsToUpdate) > 0:
eventsToUpdateList.append(eventsToUpdate)
if len(filesToStageOut) > 0:
filesToStageOutList[
workSpec.workerID] = filesToStageOut
# apfmon status update
if apfmon_status_updates and newStatus != oldStatus:
tmpQueLog.debug(
'apfmon_status_updates: {0} newStatus: {1} monStatus: {2} oldStatus: {3} workSpecStatus: {4}'
.format(apfmon_status_updates, newStatus,
monStatus, oldStatus, workSpec.status))
self.apfmon.update_worker(workSpec, monStatus)
# lock workers for fifo
if from_fifo:
# collect some attributes to be updated when workers are locked
worker_id_list = dict()
for workSpec, isChecked in zip(workSpecs, isCheckedList):
attrs = dict()
if isChecked:
attrs['checkTime'] = workSpec.checkTime
workSpec.force_not_update('checkTime')
if workSpec.has_updated_attributes():
attrs['lockedBy'] = lockedBy
workSpec.lockedBy = lockedBy
workSpec.force_not_update('lockedBy')
else:
attrs['lockedBy'] = None
worker_id_list[workSpec.workerID] = attrs
temRetLockWorker = self.dbProxy.lock_workers(
worker_id_list, harvester_config.monitor.lockInterval)
# skip if not locked
if not temRetLockWorker:
continue
# update jobs and workers
if jobSpecs is not None and len(jobSpecs) > 0:
tmpQueLog.debug(
'updating {0} jobs with {1} workers'.format(
len(jobSpecs), len(workSpecs)))
core_utils.update_job_attributes_with_workers(
mapType, jobSpecs, workSpecs, filesToStageOutList,
eventsToUpdateList)
# update local database
tmpRet = self.dbProxy.update_jobs_workers(
jobSpecs, workSpecs, lockedBy, pandaIDsList)
if not tmpRet:
for workSpec in workSpecs:
tmpLog = self.make_logger(_logger,
'id={0} workerID={1}'.format(
lockedBy,
workSpec.workerID),
method_name='run')
if from_fifo:
tmpLog.info(
'failed to update the DB. Maybe locked by other thread running with DB'
)
else:
if workSpec.status in [
WorkSpec.ST_finished, WorkSpec.ST_failed,
WorkSpec.ST_cancelled, WorkSpec.ST_missed
]:
tmpLog.info(
'worker already in final status. Skipped')
else:
tmpLog.error(
'failed to update the DB. lockInterval may be too short'
)
else:
if jobSpecs is not None:
for jobSpec in jobSpecs:
tmpLog = self.make_logger(
_logger,
'id={0} PandaID={1}'.format(
lockedBy, jobSpec.PandaID),
method_name='run')
tmpLog.debug(
'new status={0} subStatus={1} status_in_metadata={2}'
.format(
jobSpec.status, jobSpec.subStatus,
jobSpec.get_job_status_from_attributes()))
# send ACK to workers for events and files
if len(eventsToUpdateList) > 0 or len(filesToStageOutList) > 0:
for workSpec in workSpecs:
try:
messenger.acknowledge_events_files(workSpec)
except Exception:
core_utils.dump_error_message(tmpQueLog)
tmpQueLog.error(
'failed to send ACK to workerID={0}'.format(
workSpec.workerID))
# active workers for fifo
if self.monitor_fifo.enabled and workSpecs:
workSpec = workSpecs[0]
tmpOut = tmpRetMap[workSpec.workerID]
newStatus = tmpOut['newStatus']
monStatus = tmpOut['monStatus']
if newStatus in [WorkSpec.ST_submitted, WorkSpec.ST_running, WorkSpec.ST_idle] \
and workSpec.mapType != WorkSpec.MT_MultiWorkers \
and workSpec.workAttributes is not None:
timeNow = datetime.datetime.utcnow()
timeNow_timestamp = time.time()
# get lastCheckAt
_bool, lastCheckAt = workSpec.get_work_params(
'lastCheckAt')
try:
last_check_period = timeNow_timestamp - lastCheckAt
except TypeError:
last_check_period = forceEnqueueInterval + 1.0
# get lastForceEnqueueAt
_bool, lastForceEnqueueAt = workSpec.get_work_params(
'lastForceEnqueueAt')
if not (_bool and lastForceEnqueueAt is not None):
lastForceEnqueueAt = 0
# notification
intolerable_delay = max(
forceEnqueueInterval * 2,
harvester_config.monitor.checkInterval * 4)
if _bool and lastCheckAt is not None and last_check_period > harvester_config.monitor.checkInterval \
and timeNow_timestamp - harvester_config.monitor.checkInterval > self.startTimestamp:
if last_check_period > intolerable_delay:
tmpQueLog.error(
'last check period of workerID={0} is {1} sec, intolerably longer than monitor checkInterval. Will NOT enquque worker by force. Please check why monitor checks worker slowly'
.format(workSpec.workerID,
last_check_period))
else:
tmpQueLog.warning(
'last check period of workerID={0} is {1} sec, longer than monitor checkInterval'
.format(workSpec.workerID,
last_check_period))
# prepartion to enqueue fifo
if (from_fifo) \
or (not from_fifo
and timeNow_timestamp - harvester_config.monitor.sleepTime > self.startTimestamp
and last_check_period > forceEnqueueInterval
and last_check_period < intolerable_delay
and timeNow_timestamp - lastForceEnqueueAt > 86400 + forceEnqueueInterval):
if not from_fifo:
# in DB cycle
tmpQueLog.warning(
'last check period of workerID={0} is {1} sec, longer than monitor forceEnqueueInterval. Enqueue the worker by force'
.format(workSpec.workerID,
last_check_period))
workSpec.set_work_params(
{'lastForceEnqueueAt': timeNow_timestamp})
workSpec.set_work_params(
{'lastCheckAt': timeNow_timestamp})
workSpec.lockedBy = None
workSpec.force_update('lockedBy')
if monStatus in [
WorkSpec.ST_finished, WorkSpec.ST_failed,
WorkSpec.ST_cancelled
]:
# for post-processing
_bool, startFifoPreemptAt = workSpec.get_work_params(
'startFifoPreemptAt')
if not _bool or startFifoPreemptAt is None:
startFifoPreemptAt = timeNow_timestamp
workSpec.set_work_params({
'startFifoPreemptAt':
startFifoPreemptAt
})
tmpQueLog.debug(
'workerID={0} , startFifoPreemptAt: {1}'.
format(workSpec.workerID,
startFifoPreemptAt))
if timeNow_timestamp - startFifoPreemptAt < fifoMaxPreemptInterval:
workSpecsToEnqueueToHead_dict[
workSpec.workerID] = workSpecs
else:
workSpec.set_work_params({
'startFifoPreemptAt':
timeNow_timestamp
})
workSpec.modificationTime = timeNow
workSpec.force_update('modificationTime')
workSpecsToEnqueue_dict[
workSpec.workerID] = workSpecs
else:
workSpec.modificationTime = timeNow
workSpec.force_update('modificationTime')
workSpecsToEnqueue_dict[
workSpec.workerID] = workSpecs
else:
tmpQueLog.error('failed to check workers')
workSpecsToEnqueue = list(workSpecsToEnqueue_dict.values())
workSpecsToEnqueueToHead = list(workSpecsToEnqueueToHead_dict.values())
retVal = workSpecsToEnqueue, workSpecsToEnqueueToHead, timeNow_timestamp, fifoCheckInterval
tmpQueLog.debug('done')
return retVal
# wrapper for checkWorkers
def check_workers(self, mon_core, messenger, all_workers, queue_config,
tmp_log, from_fifo):
# check timeout value
try:
checkTimeout = mon_core.checkTimeout
except Exception:
try:
checkTimeout = harvester_config.monitor.checkTimeout
except Exception:
checkTimeout = None
try:
workerQueueTimeLimit = harvester_config.monitor.workerQueueTimeLimit
except AttributeError:
workerQueueTimeLimit = 172800
workersToCheck = []
thingsToPostProcess = []
retMap = dict()
for workSpec in all_workers:
eventsRequestParams = {}
eventsToUpdate = []
pandaIDs = []
workStatus = None
workAttributes = None
filesToStageOut = []
nJobsToReFill = None
if workSpec.has_work_params('finalMonStatus'):
# to post-process
_bool, finalMonStatus = workSpec.get_work_params(
'finalMonStatus')
_thing = (workSpec, (finalMonStatus, ''))
thingsToPostProcess.append(_thing)
else:
# job-level late binding
if workSpec.hasJob == 0 and workSpec.mapType != WorkSpec.MT_NoJob:
# check if job is requested
jobRequested = messenger.job_requested(workSpec)
if jobRequested:
# set ready when job is requested
workStatus = WorkSpec.ST_ready
else:
workStatus = workSpec.status
elif workSpec.nJobsToReFill in [0, None]:
# check if job is requested to refill free slots
jobRequested = messenger.job_requested(workSpec)
if jobRequested:
nJobsToReFill = jobRequested
workersToCheck.append(workSpec)
else:
workersToCheck.append(workSpec)
# add
retMap[workSpec.workerID] = {
'oldStatus': workSpec.status,
'newStatus': workStatus,
'monStatus': workStatus,
'workAttributes': workAttributes,
'filesToStageOut': filesToStageOut,
'eventsRequestParams': eventsRequestParams,
'eventsToUpdate': eventsToUpdate,
'diagMessage': '',
'pandaIDs': pandaIDs,
'nJobsToReFill': nJobsToReFill,
'isChecked': True
}
# check workers
tmp_log.debug('checking workers with plugin')
try:
if workersToCheck:
tmpStat, tmpOut = mon_core.check_workers(workersToCheck)
if not tmpStat:
tmp_log.error(
'failed to check workers with: {0}'.format(tmpOut))
workersToCheck = []
tmpOut = []
else:
tmp_log.debug('checked')
else:
tmp_log.debug('Nothing to be checked with plugin')
tmpOut = []
timeNow = datetime.datetime.utcnow()
for workSpec, (newStatus, diagMessage) in itertools.chain(
zip(workersToCheck, tmpOut), thingsToPostProcess):
workerID = workSpec.workerID
tmp_log.debug('Going to check workerID={0}'.format(workerID))
pandaIDs = []
if workerID in retMap:
# failed to check status
if newStatus is None:
tmp_log.warning(
'Failed to check workerID={0} with {1}'.format(
workerID, diagMessage))
retMap[workerID]['isChecked'] = False
# set status
if workSpec.checkTime is not None and checkTimeout is not None and \
timeNow - workSpec.checkTime > datetime.timedelta(seconds=checkTimeout):
# kill due to timeout
tmp_log.debug(
'kill workerID={0} due to consecutive check failures'
.format(workerID))
self.dbProxy.kill_worker(workSpec.workerID)
newStatus = WorkSpec.ST_cancelled
diagMessage = 'Killed by Harvester due to consecutive worker check failures. ' + diagMessage
workSpec.set_pilot_error(
PilotErrors.ERR_FAILEDBYSERVER, diagMessage)
else:
# use original status
newStatus = workSpec.status
# request kill
if messenger.kill_requested(workSpec):
tmp_log.debug(
'kill workerID={0} as requested'.format(workerID))
self.dbProxy.kill_worker(workSpec.workerID)
# stuck queuing for too long
if workSpec.status == WorkSpec.ST_submitted \
and timeNow > workSpec.submitTime + datetime.timedelta(seconds=workerQueueTimeLimit):
tmp_log.debug(
'kill workerID={0} due to queuing longer than {1} seconds'
.format(workerID, workerQueueTimeLimit))
self.dbProxy.kill_worker(workSpec.workerID)
diagMessage = 'Killed by Harvester due to worker queuing too long' + diagMessage
workSpec.set_pilot_error(
PilotErrors.ERR_FAILEDBYSERVER, diagMessage)
# expired heartbeat - only when requested in the configuration
try:
# check if the queue configuration requires checking for worker heartbeat
worker_heartbeat_limit = int(
queue_config.messenger['worker_heartbeat'])
except (AttributeError, KeyError):
worker_heartbeat_limit = None
tmp_log.debug(
'workerID={0} heartbeat limit is configured to {1}'.
format(workerID, worker_heartbeat_limit))
if worker_heartbeat_limit:
if messenger.is_alive(workSpec,
worker_heartbeat_limit):
tmp_log.debug(
'heartbeat for workerID={0} is valid'.format(
workerID))
else:
tmp_log.debug(
'heartbeat for workerID={0} expired: sending kill request'
.format(workerID))
self.dbProxy.kill_worker(workSpec.workerID)
diagMessage = 'Killed by Harvester due to worker heartbeat expired. ' + diagMessage
workSpec.set_pilot_error(
PilotErrors.ERR_FAILEDBYSERVER, diagMessage)
# get work attributes
workAttributes = messenger.get_work_attributes(workSpec)
retMap[workerID]['workAttributes'] = workAttributes
# get output files
filesToStageOut = messenger.get_files_to_stage_out(
workSpec)
retMap[workerID]['filesToStageOut'] = filesToStageOut
# get events to update
if workSpec.eventsRequest in [
WorkSpec.EV_useEvents, WorkSpec.EV_requestEvents
]:
eventsToUpdate = messenger.events_to_update(workSpec)
retMap[workerID]['eventsToUpdate'] = eventsToUpdate
# request events
if workSpec.eventsRequest == WorkSpec.EV_useEvents:
eventsRequestParams = messenger.events_requested(
workSpec)
retMap[workerID][
'eventsRequestParams'] = eventsRequestParams
# get PandaIDs for pull model
if workSpec.mapType == WorkSpec.MT_NoJob:
pandaIDs = messenger.get_panda_ids(workSpec)
retMap[workerID]['pandaIDs'] = pandaIDs
# keep original new status
retMap[workerID]['monStatus'] = newStatus
# set running or idle while there are events to update or files to stage out
if newStatus in [
WorkSpec.ST_finished, WorkSpec.ST_failed,
WorkSpec.ST_cancelled
]:
if len(retMap[workerID]['filesToStageOut']) > 0 or \
len(retMap[workerID]['eventsToUpdate']) > 0:
if workSpec.status == WorkSpec.ST_running:
newStatus = WorkSpec.ST_running
else:
newStatus = WorkSpec.ST_idle
elif not workSpec.is_post_processed():
if not queue_config.is_no_heartbeat_status(
newStatus):
# post processing unless heartbeat is suppressed
jobSpecs = self.dbProxy.get_jobs_with_worker_id(
workSpec.workerID,
None,
True,
only_running=True,
slim=True)
# post processing
messenger.post_processing(
workSpec, jobSpecs, workSpec.mapType)
workSpec.post_processed()
if workSpec.status == WorkSpec.ST_running:
newStatus = WorkSpec.ST_running
else:
newStatus = WorkSpec.ST_idle
# reset modification time to immediately trigger subsequent lookup
if not self.monitor_fifo.enabled:
workSpec.trigger_next_lookup()
retMap[workerID]['newStatus'] = newStatus
retMap[workerID]['diagMessage'] = diagMessage
else:
tmp_log.debug(
'workerID={0} not in retMap'.format(workerID))
return True, retMap
except Exception:
core_utils.dump_error_message(tmp_log)
return False, None
def start(self):
# thread list
thrList = []
# Credential Manager
from pandaharvester.harvesterbody.cred_manager import CredManager
thr = CredManager(single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.execute()
thr.start()
thrList.append(thr)
# Command manager
from pandaharvester.harvesterbody.command_manager import CommandManager
thr = CommandManager(self.communicatorPool, self.queueConfigMapper, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Cacher
from pandaharvester.harvesterbody.cacher import Cacher
thr = Cacher(self.communicatorPool, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.execute(force_update=True, skip_lock=True)
thr.start()
thrList.append(thr)
# Watcher
from pandaharvester.harvesterbody.watcher import Watcher
thr = Watcher(single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Job Fetcher
from pandaharvester.harvesterbody.job_fetcher import JobFetcher
nThr = harvester_config.jobfetcher.nThreads
for iThr in range(nThr):
thr = JobFetcher(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Propagator
from pandaharvester.harvesterbody.propagator import Propagator
nThr = harvester_config.propagator.nThreads
for iThr in range(nThr):
thr = Propagator(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Monitor
from pandaharvester.harvesterbody.monitor import Monitor
nThr = harvester_config.monitor.nThreads
for iThr in range(nThr):
thr = Monitor(self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Preparator
from pandaharvester.harvesterbody.preparator import Preparator
nThr = harvester_config.preparator.nThreads
for iThr in range(nThr):
thr = Preparator(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Submitter
from pandaharvester.harvesterbody.submitter import Submitter
nThr = harvester_config.submitter.nThreads
for iThr in range(nThr):
thr = Submitter(self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Stager
from pandaharvester.harvesterbody.stager import Stager
nThr = harvester_config.stager.nThreads
for iThr in range(nThr):
thr = Stager(self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# EventFeeder
from pandaharvester.harvesterbody.event_feeder import EventFeeder
nThr = harvester_config.eventfeeder.nThreads
for iThr in range(nThr):
thr = EventFeeder(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Sweeper
from pandaharvester.harvesterbody.sweeper import Sweeper
nThr = harvester_config.sweeper.nThreads
for iThr in range(nThr):
thr = Sweeper(self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Service monitor
try:
sm_active = harvester_config.service_monitor.active
except:
sm_active = False
if sm_active:
from pandaharvester.harvesterbody.service_monitor import ServiceMonitor
thr = ServiceMonitor(options.pid, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Report itself to APF Mon
apf_mon = Apfmon(self.queueConfigMapper)
apf_mon.create_factory()
apf_mon.create_labels()
##################
# loop on stop event to be interruptable since thr.join blocks signal capture in python 2.7
while True:
if self.singleMode or not self.daemonMode:
break
self.stopEvent.wait(1)
if self.stopEvent.is_set():
break
##################
# join
if self.daemonMode:
for thr in thrList:
thr.join()
class WorkerAdjuster(object):
# constructor
def __init__(self, queue_config_mapper):
self.queueConfigMapper = queue_config_mapper
self.pluginFactory = PluginFactory()
self.dbProxy = DBProxy()
self.throttlerMap = dict()
self.apf_mon = Apfmon(self.queueConfigMapper)
try:
self.maxNewWorkers = harvester_config.submitter.maxNewWorkers
except AttributeError:
self.maxNewWorkers = None
# define number of workers to submit based on various information
def define_num_workers(self, static_num_workers, site_name):
tmpLog = core_utils.make_logger(_logger, 'site={0}'.format(site_name), method_name='define_num_workers')
tmpLog.debug('start')
tmpLog.debug('static_num_workers: {0}'.format(static_num_workers))
dyn_num_workers = copy.deepcopy(static_num_workers)
try:
# get queue status
queueStat = self.dbProxy.get_cache("panda_queues.json", None)
if queueStat is None:
queueStat = dict()
else:
queueStat = queueStat.data
# get job statistics
job_stats = self.dbProxy.get_cache("job_statistics.json", None)
if job_stats is None:
job_stats = dict()
else:
job_stats = job_stats.data
# define num of new workers
for queueName in static_num_workers:
# get queue
queueConfig = self.queueConfigMapper.get_queue(queueName)
workerLimits_dict = self.dbProxy.get_worker_limits(queueName)
maxWorkers = workerLimits_dict.get('maxWorkers', 0)
nQueueLimit = workerLimits_dict.get('nQueueLimitWorker', 0)
nQueueLimitPerRT = workerLimits_dict['nQueueLimitWorkerPerRT']
nQueue_total, nReady_total, nRunning_total = 0, 0, 0
apf_msg = None
apf_data = None
for resource_type, tmpVal in iteritems(static_num_workers[queueName]):
tmpLog.debug('Processing queue {0} resource {1} with static_num_workers {2}'.
format(queueName, resource_type, tmpVal))
# set 0 to num of new workers when the queue is disabled
if queueName in queueStat and queueStat[queueName]['status'] in ['offline', 'standby',
'maintenance']:
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = 0
retMsg = 'set nNewWorkers=0 since status={0}'.format(queueStat[queueName]['status'])
tmpLog.debug(retMsg)
apf_msg = 'Not submitting workers since queue status = {0}'.format(queueStat[queueName]['status'])
continue
# protection against not-up-to-date queue config
if queueConfig is None:
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = 0
retMsg = 'set nNewWorkers=0 due to missing queueConfig'
tmpLog.debug(retMsg)
apf_msg = 'Not submitting workers because of missing queueConfig'
continue
# get throttler
if queueName not in self.throttlerMap:
if hasattr(queueConfig, 'throttler'):
throttler = self.pluginFactory.get_plugin(queueConfig.throttler)
else:
throttler = None
self.throttlerMap[queueName] = throttler
# check throttler
throttler = self.throttlerMap[queueName]
if throttler is not None:
toThrottle, tmpMsg = throttler.to_be_throttled(queueConfig)
if toThrottle:
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = 0
retMsg = 'set nNewWorkers=0 by {0}:{1}'.format(throttler.__class__.__name__, tmpMsg)
tmpLog.debug(retMsg)
continue
# check stats
nQueue = tmpVal['nQueue']
nReady = tmpVal['nReady']
nRunning = tmpVal['nRunning']
if resource_type != 'ANY':
nQueue_total += nQueue
nReady_total += nReady
nRunning_total += nRunning
if queueConfig.runMode == 'slave':
nNewWorkersDef = tmpVal['nNewWorkers']
if nNewWorkersDef == 0:
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = 0
retMsg = 'set nNewWorkers=0 by panda in slave mode'
tmpLog.debug(retMsg)
continue
else:
nNewWorkersDef = None
# define num of new workers based on static site config
nNewWorkers = 0
if nQueue >= nQueueLimitPerRT > 0:
# enough queued workers
retMsg = 'No nNewWorkers since nQueue({0})>=nQueueLimitPerRT({1})'.format(nQueue, nQueueLimitPerRT)
tmpLog.debug(retMsg)
pass
elif (nQueue + nReady + nRunning) >= maxWorkers > 0:
# enough workers in the system
retMsg = 'No nNewWorkers since nQueue({0}) + nReady({1}) + nRunning({2}) '.format(nQueue,
nReady,
nRunning)
retMsg += '>= maxWorkers({0})'.format(maxWorkers)
tmpLog.debug(retMsg)
pass
else:
maxQueuedWorkers = None
if nQueueLimitPerRT > 0: # there is a limit set for the queue
maxQueuedWorkers = nQueueLimitPerRT
# Reset the maxQueueWorkers according to particular
if nNewWorkersDef is not None: # don't surpass limits given centrally
maxQueuedWorkers_slave = nNewWorkersDef + nQueue
if maxQueuedWorkers is not None:
maxQueuedWorkers = min(maxQueuedWorkers_slave, maxQueuedWorkers)
else:
maxQueuedWorkers = maxQueuedWorkers_slave
elif queueConfig.mapType == 'NoJob': # for pull mode, limit to activated jobs
# limit the queue to the number of activated jobs to avoid empty pilots
try:
n_activated = max(job_stats[queueName]['activated'], 1) # avoid no activity queues
queue_limit = maxQueuedWorkers
maxQueuedWorkers = min(n_activated, maxQueuedWorkers)
tmpLog.debug('limiting maxQueuedWorkers to min(n_activated={0}, queue_limit={1})'.
format(n_activated, queue_limit))
except KeyError:
tmpLog.warning('n_activated not defined, defaulting to configured queue limits')
pass
if maxQueuedWorkers is None: # no value found, use default value
maxQueuedWorkers = 1
# new workers
nNewWorkers = max(maxQueuedWorkers - nQueue, 0)
tmpLog.debug('setting nNewWorkers to {0} in maxQueuedWorkers calculation'
.format(nNewWorkers))
if maxWorkers > 0:
nNewWorkers = min(nNewWorkers, max(maxWorkers - nQueue - nReady - nRunning, 0))
tmpLog.debug('setting nNewWorkers to {0} to respect maxWorkers'
.format(nNewWorkers))
if queueConfig.maxNewWorkersPerCycle > 0:
nNewWorkers = min(nNewWorkers, queueConfig.maxNewWorkersPerCycle)
tmpLog.debug('setting nNewWorkers to {0} in order to respect maxNewWorkersPerCycle'
.format(nNewWorkers))
if self.maxNewWorkers is not None and self.maxNewWorkers > 0:
nNewWorkers = min(nNewWorkers, self.maxNewWorkers)
tmpLog.debug('setting nNewWorkers to {0} in order to respect universal maxNewWorkers'
.format(nNewWorkers))
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = nNewWorkers
# adjust nNewWorkers for UCORE to let aggregations over RT respect nQueueLimitWorker and maxWorkers
if queueConfig is None:
maxNewWorkersPerCycle = 0
retMsg = 'set maxNewWorkersPerCycle=0 in UCORE aggregation due to missing queueConfig'
tmpLog.debug(retMsg)
else:
maxNewWorkersPerCycle = queueConfig.maxNewWorkersPerCycle
if len(dyn_num_workers[queueName]) > 1:
total_new_workers_rts = sum( dyn_num_workers[queueName][_rt]['nNewWorkers']
if _rt != 'ANY' else 0
for _rt in dyn_num_workers[queueName] )
nNewWorkers_max_agg = min(
max(nQueueLimit - nQueue_total, 0),
max(maxWorkers - nQueue_total - nReady_total - nRunning_total, 0),
)
if maxNewWorkersPerCycle >= 0:
nNewWorkers_max_agg = min(nNewWorkers_max_agg, maxNewWorkersPerCycle)
if self.maxNewWorkers is not None and self.maxNewWorkers > 0:
nNewWorkers_max_agg = min(nNewWorkers_max_agg, self.maxNewWorkers)
# exceeded max, to adjust
if total_new_workers_rts > nNewWorkers_max_agg:
if nNewWorkers_max_agg == 0:
for resource_type in dyn_num_workers[queueName]:
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = 0
tmpLog.debug('No nNewWorkers since nNewWorkers_max_agg=0 for UCORE')
else:
tmpLog.debug('nNewWorkers_max_agg={0} for UCORE'.format(nNewWorkers_max_agg))
_d = dyn_num_workers[queueName].copy()
del _d['ANY']
simple_rt_nw_list = [ [_rt, _d[_rt].get('nNewWorkers', 0), 0] for _rt in _d ]
_countdown = nNewWorkers_max_agg
for _rt_list in simple_rt_nw_list:
resource_type, nNewWorkers_orig, _r = _rt_list
nNewWorkers, remainder = divmod(nNewWorkers_orig*nNewWorkers_max_agg, total_new_workers_rts)
dyn_num_workers[queueName][resource_type]['nNewWorkers'] = nNewWorkers
_rt_list[2] = remainder
_countdown -= nNewWorkers
_s_list = sorted(simple_rt_nw_list, key=(lambda x: x[1]))
sorted_rt_nw_list = sorted(_s_list, key=(lambda x: x[2]), reverse=True)
for resource_type, nNewWorkers_orig, remainder in sorted_rt_nw_list:
if _countdown <= 0:
break
dyn_num_workers[queueName][resource_type]['nNewWorkers'] += 1
_countdown -= 1
for resource_type in dyn_num_workers[queueName]:
if resource_type == 'ANY':
continue
nNewWorkers = dyn_num_workers[queueName][resource_type]['nNewWorkers']
tmpLog.debug('setting nNewWorkers to {0} of type {1} in order to respect RT aggregations for UCORE'
.format(nNewWorkers, resource_type))
if not apf_msg:
apf_data = copy.deepcopy(dyn_num_workers[queueName])
self.apf_mon.update_label(queueName, apf_msg, apf_data)
# dump
tmpLog.debug('defined {0}'.format(str(dyn_num_workers)))
return dyn_num_workers
except Exception:
# dump error
errMsg = core_utils.dump_error_message(tmpLog)
return None
def start(self):
# thread list
thrList = []
# Credential Manager
from pandaharvester.harvesterbody.cred_manager import CredManager
thr = CredManager(self.queueConfigMapper, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.execute()
thr.start()
thrList.append(thr)
# Command manager
from pandaharvester.harvesterbody.command_manager import CommandManager
thr = CommandManager(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Cacher
from pandaharvester.harvesterbody.cacher import Cacher
thr = Cacher(self.communicatorPool, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.execute(force_update=True, skip_lock=True)
thr.start()
thrList.append(thr)
# Watcher
from pandaharvester.harvesterbody.watcher import Watcher
thr = Watcher(single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Job Fetcher
from pandaharvester.harvesterbody.job_fetcher import JobFetcher
nThr = harvester_config.jobfetcher.nThreads
for iThr in range(nThr):
thr = JobFetcher(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Propagator
from pandaharvester.harvesterbody.propagator import Propagator
nThr = harvester_config.propagator.nThreads
for iThr in range(nThr):
thr = Propagator(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Monitor
from pandaharvester.harvesterbody.monitor import Monitor
nThr = harvester_config.monitor.nThreads
for iThr in range(nThr):
thr = Monitor(self.queueConfigMapper, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Preparator
from pandaharvester.harvesterbody.preparator import Preparator
nThr = harvester_config.preparator.nThreads
for iThr in range(nThr):
thr = Preparator(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Submitter
from pandaharvester.harvesterbody.submitter import Submitter
nThr = harvester_config.submitter.nThreads
for iThr in range(nThr):
thr = Submitter(self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Stager
from pandaharvester.harvesterbody.stager import Stager
nThr = harvester_config.stager.nThreads
for iThr in range(nThr):
thr = Stager(self.queueConfigMapper, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# EventFeeder
from pandaharvester.harvesterbody.event_feeder import EventFeeder
nThr = harvester_config.eventfeeder.nThreads
for iThr in range(nThr):
thr = EventFeeder(self.communicatorPool,
self.queueConfigMapper,
single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Sweeper
from pandaharvester.harvesterbody.sweeper import Sweeper
nThr = harvester_config.sweeper.nThreads
for iThr in range(nThr):
thr = Sweeper(self.queueConfigMapper, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Service monitor
try:
sm_active = harvester_config.service_monitor.active
except Exception:
sm_active = False
if sm_active:
from pandaharvester.harvesterbody.service_monitor import ServiceMonitor
thr = ServiceMonitor(options.pid, single_mode=self.singleMode)
thr.set_stop_event(self.stopEvent)
thr.start()
thrList.append(thr)
# Report itself to APF Mon
apf_mon = Apfmon(self.queueConfigMapper)
apf_mon.create_factory()
apf_mon.create_labels()
##################
# loop on stop event to be interruptable since thr.join blocks signal capture in python 2.7
while True:
if self.singleMode or not self.daemonMode:
break
self.stopEvent.wait(1)
if self.stopEvent.is_set():
break
##################
# join
if self.daemonMode:
for thr in thrList:
thr.join()
class Submitter(AgentBase):
# constructor
def __init__(self, queue_config_mapper, single_mode=False):
AgentBase.__init__(self, single_mode)
self.queueConfigMapper = queue_config_mapper
self.dbProxy = DBProxy()
self.workerMaker = WorkerMaker()
self.workerAdjuster = WorkerAdjuster(queue_config_mapper)
self.pluginFactory = PluginFactory()
self.monitor_fifo = MonitorFIFO()
self.apfmon = Apfmon(self.queueConfigMapper)
# main loop
def run(self):
lockedBy = 'submitter-{0}'.format(self.get_pid())
monitor_fifo = self.monitor_fifo
queueLockInterval = getattr(harvester_config.submitter, 'queueLockInterval',
harvester_config.submitter.lockInterval)
while True:
sw_main = core_utils.get_stopwatch()
mainLog = self.make_logger(_logger, 'id={0}'.format(lockedBy), method_name='run')
mainLog.debug('getting queues to submit workers')
# get queues associated to a site to submit workers
curWorkers, siteName, resMap = self.dbProxy.get_queues_to_submit(harvester_config.submitter.nQueues,
harvester_config.submitter.lookupTime,
harvester_config.submitter.lockInterval,
lockedBy, queueLockInterval)
submitted = False
if siteName is not None:
mainLog.debug('got {0} queues for site {1}'.format(len(curWorkers), siteName))
# get commands
comStr = '{0}:{1}'.format(CommandSpec.COM_setNWorkers, siteName)
commandSpecs = self.dbProxy.get_commands_for_receiver('submitter', comStr)
mainLog.debug('got {0} {1} commands'.format(commandSpecs, comStr))
for commandSpec in commandSpecs:
newLimits = self.dbProxy.set_queue_limit(siteName, commandSpec.params)
for tmpResource, tmpNewVal in iteritems(newLimits):
# if available, overwrite new worker value with the command from panda server
if tmpResource in resMap:
tmpQueueName = resMap[tmpResource]
if tmpQueueName in curWorkers:
curWorkers[tmpQueueName][tmpResource]['nNewWorkers'] = tmpNewVal
# define number of new workers
if len(curWorkers) == 0:
n_workers_per_queue_and_rt = dict()
else:
n_workers_per_queue_and_rt = self.workerAdjuster.define_num_workers(curWorkers, siteName)
if n_workers_per_queue_and_rt is None:
mainLog.error('WorkerAdjuster failed to define the number of workers')
elif len(n_workers_per_queue_and_rt) == 0:
pass
else:
# loop over all queues and resource types
for queueName in n_workers_per_queue_and_rt:
for resource_type, tmpVal in iteritems(n_workers_per_queue_and_rt[queueName]):
tmpLog = self.make_logger(_logger, 'id={0} queue={1} rtype={2}'.format(lockedBy,
queueName,
resource_type),
method_name='run')
try:
tmpLog.debug('start')
tmpLog.debug('workers status: %s' % tmpVal)
nWorkers = tmpVal['nNewWorkers'] + tmpVal['nReady']
nReady = tmpVal['nReady']
# check queue
if not self.queueConfigMapper.has_queue(queueName):
tmpLog.error('config not found')
continue
# no new workers
if nWorkers == 0:
tmpLog.debug('skipped since no new worker is needed based on current stats')
continue
# get queue
queueConfig = self.queueConfigMapper.get_queue(queueName)
workerMakerCore = self.workerMaker.get_plugin(queueConfig)
# check if resource is ready
if hasattr(workerMakerCore, 'dynamicSizing') and workerMakerCore.dynamicSizing is True:
numReadyResources = self.workerMaker.num_ready_resources(queueConfig,
resource_type,
workerMakerCore)
tmpLog.debug('numReadyResources: %s' % numReadyResources)
if not numReadyResources:
if hasattr(workerMakerCore, 'staticWorkers'):
nQRWorkers = tmpVal['nQueue'] + tmpVal['nRunning']
tmpLog.debug('staticWorkers: %s, nQRWorkers(Queue+Running): %s' %
(workerMakerCore.staticWorkers, nQRWorkers))
if nQRWorkers >= workerMakerCore.staticWorkers:
tmpLog.debug('No left static workers, skip')
continue
else:
nWorkers = min(workerMakerCore.staticWorkers - nQRWorkers, nWorkers)
tmpLog.debug('staticWorkers: %s, nWorkers: %s' %
(workerMakerCore.staticWorkers, nWorkers))
else:
tmpLog.debug('skip since no resources are ready')
continue
else:
nWorkers = min(nWorkers, numReadyResources)
# post action of worker maker
if hasattr(workerMakerCore, 'skipOnFail') and workerMakerCore.skipOnFail is True:
skipOnFail = True
else:
skipOnFail = False
# actions based on mapping type
if queueConfig.mapType == WorkSpec.MT_NoJob:
# workers without jobs
jobChunks = []
for i in range(nWorkers):
jobChunks.append([])
elif queueConfig.mapType == WorkSpec.MT_OneToOne:
# one worker per one job
jobChunks = self.dbProxy.get_job_chunks_for_workers(
queueName,
nWorkers, nReady, 1, None,
queueConfig.useJobLateBinding,
harvester_config.submitter.checkInterval,
harvester_config.submitter.lockInterval,
lockedBy)
elif queueConfig.mapType == WorkSpec.MT_MultiJobs:
# one worker for multiple jobs
nJobsPerWorker = self.workerMaker.get_num_jobs_per_worker(queueConfig,
nWorkers,
resource_type,
maker=workerMakerCore)
tmpLog.debug('nJobsPerWorker={0}'.format(nJobsPerWorker))
jobChunks = self.dbProxy.get_job_chunks_for_workers(
queueName,
nWorkers, nReady, nJobsPerWorker, None,
queueConfig.useJobLateBinding,
harvester_config.submitter.checkInterval,
harvester_config.submitter.lockInterval,
lockedBy,
queueConfig.allowJobMixture)
elif queueConfig.mapType == WorkSpec.MT_MultiWorkers:
# multiple workers for one job
nWorkersPerJob = self.workerMaker.get_num_workers_per_job(queueConfig,
nWorkers,
resource_type,
maker=workerMakerCore)
maxWorkersPerJob = self.workerMaker.get_max_workers_per_job_in_total(
queueConfig, resource_type, maker=workerMakerCore)
maxWorkersPerJobPerCycle = self.workerMaker.get_max_workers_per_job_per_cycle(
queueConfig, resource_type, maker=workerMakerCore)
tmpLog.debug('nWorkersPerJob={0}'.format(nWorkersPerJob))
jobChunks = self.dbProxy.get_job_chunks_for_workers(
queueName,
nWorkers, nReady, None, nWorkersPerJob,
queueConfig.useJobLateBinding,
harvester_config.submitter.checkInterval,
harvester_config.submitter.lockInterval,
lockedBy, max_workers_per_job_in_total=maxWorkersPerJob,
max_workers_per_job_per_cycle=maxWorkersPerJobPerCycle)
else:
tmpLog.error('unknown mapType={0}'.format(queueConfig.mapType))
continue
tmpLog.debug('got {0} job chunks'.format(len(jobChunks)))
if len(jobChunks) == 0:
continue
# make workers
okChunks, ngChunks = self.workerMaker.make_workers(jobChunks, queueConfig,
nReady, resource_type,
maker=workerMakerCore)
if len(ngChunks) == 0:
tmpLog.debug('successfully made {0} workers'.format(len(okChunks)))
else:
tmpLog.debug('made {0} workers, while {1} workers failed'.format(len(okChunks),
len(ngChunks)))
timeNow = datetime.datetime.utcnow()
timeNow_timestamp = time.time()
pandaIDs = set()
# NG (=not good)
for ngJobs in ngChunks:
for jobSpec in ngJobs:
if skipOnFail:
# release jobs when workers are not made
pandaIDs.add(jobSpec.PandaID)
else:
jobSpec.status = 'failed'
jobSpec.subStatus = 'failed_to_make'
jobSpec.stateChangeTime = timeNow
jobSpec.lockedBy = None
errStr = 'failed to make a worker'
jobSpec.set_pilot_error(PilotErrors.ERR_SETUPFAILURE, errStr)
jobSpec.trigger_propagation()
self.dbProxy.update_job(jobSpec, {'lockedBy': lockedBy,
'subStatus': 'prepared'})
# OK
workSpecList = []
if len(okChunks) > 0:
for workSpec, okJobs in okChunks:
# has job
if (queueConfig.useJobLateBinding and workSpec.workerID is None) \
or queueConfig.mapType == WorkSpec.MT_NoJob:
workSpec.hasJob = 0
else:
workSpec.hasJob = 1
if workSpec.nJobsToReFill in [None, 0]:
workSpec.set_jobspec_list(okJobs)
else:
# refill free slots during the worker is running
workSpec.set_jobspec_list(okJobs[:workSpec.nJobsToReFill])
workSpec.nJobsToReFill = None
for jobSpec in okJobs[workSpec.nJobsToReFill:]:
pandaIDs.add(jobSpec.PandaID)
workSpec.set_num_jobs_with_list()
# map type
workSpec.mapType = queueConfig.mapType
# queue name
workSpec.computingSite = queueConfig.queueName
# set access point
workSpec.accessPoint = queueConfig.messenger['accessPoint']
# sync level
workSpec.syncLevel = queueConfig.get_synchronization_level()
# events
if len(okJobs) > 0 and \
('eventService' in okJobs[0].jobParams or
'cloneJob' in okJobs[0].jobParams):
workSpec.eventsRequest = WorkSpec.EV_useEvents
workSpecList.append(workSpec)
if len(workSpecList) > 0:
sw = core_utils.get_stopwatch()
# get plugin for submitter
submitterCore = self.pluginFactory.get_plugin(queueConfig.submitter)
if submitterCore is None:
# not found
tmpLog.error(
'submitter plugin for {0} not found'.format(jobSpec.computingSite))
continue
# get plugin for messenger
messenger = self.pluginFactory.get_plugin(queueConfig.messenger)
if messenger is None:
# not found
tmpLog.error(
'messenger plugin for {0} not found'.format(jobSpec.computingSite))
continue
# setup access points
messenger.setup_access_points(workSpecList)
# feed jobs
for workSpec in workSpecList:
if workSpec.hasJob == 1:
tmpStat = messenger.feed_jobs(workSpec, workSpec.get_jobspec_list())
if tmpStat is False:
tmpLog.error(
'failed to send jobs to workerID={0}'.format(workSpec.workerID))
else:
tmpLog.debug(
'sent jobs to workerID={0} with {1}'.format(workSpec.workerID,
tmpStat))
# insert workers
self.dbProxy.insert_workers(workSpecList, lockedBy)
# submit
sw.reset()
tmpLog.info('submitting {0} workers'.format(len(workSpecList)))
workSpecList, tmpRetList, tmpStrList = self.submit_workers(submitterCore,
workSpecList)
tmpLog.debug('done submitting {0} workers'.format(len(workSpecList))
+ sw.get_elapsed_time())
# collect successful jobs
okPandaIDs = set()
for iWorker, (tmpRet, tmpStr) in enumerate(zip(tmpRetList, tmpStrList)):
if tmpRet:
workSpec, jobList = okChunks[iWorker]
jobList = workSpec.get_jobspec_list()
if jobList is not None:
for jobSpec in jobList:
okPandaIDs.add(jobSpec.PandaID)
# loop over all workers
for iWorker, (tmpRet, tmpStr) in enumerate(zip(tmpRetList, tmpStrList)):
workSpec, jobList = okChunks[iWorker]
# set harvesterHost
workSpec.harvesterHost = socket.gethostname()
# use associated job list since it can be truncated for re-filling
jobList = workSpec.get_jobspec_list()
# set status
if not tmpRet:
# failed submission
errStr = 'failed to submit a workerID={0} with {1}'.format(
workSpec.workerID,
tmpStr)
tmpLog.error(errStr)
workSpec.set_status(WorkSpec.ST_missed)
workSpec.set_dialog_message(tmpStr)
workSpec.set_pilot_error(PilotErrors.ERR_SETUPFAILURE, errStr)
if jobList is not None:
# increment attempt number
newJobList = []
for jobSpec in jobList:
# skip if successful with another worker
if jobSpec.PandaID in okPandaIDs:
continue
if jobSpec.submissionAttempts is None:
jobSpec.submissionAttempts = 0
jobSpec.submissionAttempts += 1
# max attempt or permanent error
if tmpRet is False or \
jobSpec.submissionAttempts >= \
queueConfig.maxSubmissionAttempts:
newJobList.append(jobSpec)
else:
self.dbProxy.increment_submission_attempt(
jobSpec.PandaID,
jobSpec.submissionAttempts)
jobList = newJobList
elif queueConfig.useJobLateBinding and workSpec.hasJob == 1:
# directly go to running after feeding jobs for late biding
workSpec.set_status(WorkSpec.ST_running)
else:
# normal successful submission
workSpec.set_status(WorkSpec.ST_submitted)
workSpec.submitTime = timeNow
workSpec.modificationTime = timeNow
workSpec.checkTime = timeNow
if self.monitor_fifo.enabled:
workSpec.set_work_params({'lastCheckAt': timeNow_timestamp})
# prefetch events
if tmpRet and workSpec.hasJob == 1 and \
workSpec.eventsRequest == WorkSpec.EV_useEvents and \
queueConfig.prefetchEvents:
workSpec.eventsRequest = WorkSpec.EV_requestEvents
eventsRequestParams = dict()
for jobSpec in jobList:
eventsRequestParams[jobSpec.PandaID] = \
{'pandaID': jobSpec.PandaID,
'taskID': jobSpec.taskID,
'jobsetID': jobSpec.jobParams['jobsetID'],
'nRanges': max(int(math.ceil(workSpec.nCore / len(jobList))),
jobSpec.jobParams['coreCount']),
}
workSpec.eventsRequestParams = eventsRequestParams
# register worker
tmpStat = self.dbProxy.register_worker(workSpec, jobList, lockedBy)
if jobList is not None:
for jobSpec in jobList:
pandaIDs.add(jobSpec.PandaID)
if tmpStat:
if tmpRet:
tmpStr = \
'submitted a workerID={0} for PandaID={1} with batchID={2}'
tmpLog.info(tmpStr.format(workSpec.workerID,
jobSpec.PandaID,
workSpec.batchID))
else:
tmpStr = 'failed to submit a workerID={0} for PandaID={1}'
tmpLog.error(tmpStr.format(workSpec.workerID,
jobSpec.PandaID))
else:
tmpStr = \
'failed to register a worker for PandaID={0} with batchID={1}'
tmpLog.error(tmpStr.format(jobSpec.PandaID, workSpec.batchID))
# enqueue to monitor fifo
if self.monitor_fifo.enabled \
and queueConfig.mapType != WorkSpec.MT_MultiWorkers:
workSpecsToEnqueue = \
[[w] for w in workSpecList if w.status
in (WorkSpec.ST_submitted, WorkSpec.ST_running)]
check_delay = min(
getattr(harvester_config.monitor, 'eventBasedCheckInterval',
harvester_config.monitor.checkInterval),
getattr(harvester_config.monitor, 'fifoCheckInterval',
harvester_config.monitor.checkInterval))
monitor_fifo.put((queueName, workSpecsToEnqueue), time.time() + check_delay)
mainLog.debug('put workers to monitor FIFO')
submitted = True
# release jobs
self.dbProxy.release_jobs(pandaIDs, lockedBy)
tmpLog.info('done')
except Exception:
core_utils.dump_error_message(tmpLog)
# release the site
self.dbProxy.release_site(siteName, lockedBy)
if sw_main.get_elapsed_time_in_sec() > queueLockInterval:
mainLog.warning('a submitter cycle was longer than queueLockInterval {0} sec'.format(queueLockInterval)
+ sw_main.get_elapsed_time())
mainLog.debug('done')
# define sleep interval
if siteName is None:
sleepTime = harvester_config.submitter.sleepTime
else:
sleepTime = 0
if submitted and hasattr(harvester_config.submitter, 'minSubmissionInterval'):
interval = harvester_config.submitter.minSubmissionInterval
if interval > 0:
newTime = datetime.datetime.utcnow() + datetime.timedelta(seconds=interval)
self.dbProxy.update_panda_queue_attribute('submitTime', newTime, site_name=siteName)
# time the cycle
mainLog.debug('done a submitter cycle' + sw_main.get_elapsed_time())
# check if being terminated
if self.terminated(sleepTime):
mainLog.debug('terminated')
return
# wrapper for submitWorkers to skip ready workers
def submit_workers(self, submitter_core, workspec_list):
retList = []
strList = []
newSpecList = []
workersToSubmit = []
for workSpec in workspec_list:
if workSpec.status in [WorkSpec.ST_ready, WorkSpec.ST_running]:
newSpecList.append(workSpec)
retList.append(True)
strList.append('')
else:
workersToSubmit.append(workSpec)
tmpRetList = submitter_core.submit_workers(workersToSubmit)
# submit the workers to the monitoring
self.apfmon.create_workers(workersToSubmit)
for tmpRet, tmpStr in tmpRetList:
retList.append(tmpRet)
strList.append(tmpStr)
newSpecList += workersToSubmit
return newSpecList, retList, strList
