def run(self):
try:
# print '%s start' % self.uid
while not self.term.is_set():
# print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 4 and self.pnum == 1:
print "4"
ru.raise_on(self.uid, RAISE_ON)
# print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
except SystemExit:
print '%s exit' % (self.uid)
except KeyboardInterrupt:
print '%s intr' % (self.uid)
finally:
print '%s final' % (self.uid)
def run(self):
self.uid = self.uid + '.thread'
self.log = ru.get_logger('radical.' + self.uid, level=self.verbose)
# a simple worker routine which sleeps repeatedly for a random number of
# seconds, until a term signal is set. The given 'worker' can be a thread
# or process, or in fact anything which has a self.uid and self.term.
try:
self.log.info('%-10s : work start' % self.uid)
while not self.term.is_set():
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
self.log.info('%-10s : %ds sleep start' % (self.uid, item))
time.sleep(item)
self.log.info('%-10s : %ds sleep stop' % (self.uid, item))
ru.raise_on('work')
self.log.info('%-10s : work term requested' % self.uid)
except Exception as e:
self.log.info('%-10s : work fail [%s]' % (self.uid, e))
raise
self.log.info('%-10s : thread exit requested' % self.uid)
def run(self):
self.uid = self.uid + '.thread'
self.log = ru.get_logger('radical.' + self.uid, level=self.verbose)
# a simple worker routine which sleeps repeatedly for a random number of
# seconds, until a term signal is set. The given 'worker' can be a thread
# or process, or in fact anything which has a self.uid and self.term.
try:
self.log.info('%-10s : work start' % self.uid)
while not self.term.is_set():
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
self.log.info('%-10s : %ds sleep start' % (self.uid, item))
time.sleep(item)
self.log.info('%-10s : %ds sleep stop' % (self.uid, item))
ru.raise_on('work')
self.log.info('%-10s : work term requested' % self.uid)
except Exception as e:
self.log.info('%-10s : work fail [%s]' % (self.uid, e))
raise
self.log.info('%-10s : thread exit requested' % self.uid)
def work(self):
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
self.log.info('%-10s : %ds sleep start' % (self.uid, item))
time.sleep(item)
self.log.info('%-10s : %ds sleep stop' % (self.uid, item))
ru.raise_on('work')
def run(self):
try:
self.log.info('%-10s : start' % self.uid)
while not self.term.is_set():
time.sleep(0.1) # start things
ru.raise_on('watch')
for t in self.things:
if not t.is_alive():
self.log.info('%-10s : %s died' % (self.uid, t.uid))
# a child died. We kill the other children and
# terminate.
# self.stop()
return
self.log.info('%-10s : %s ok' % (self.uid, t.uid))
except ru.ThreadExit:
raise RuntimeError('%-10s : watcher exit requested [%s]' % \
(self.uid, self.ident))
except Exception as e:
raise RuntimeError('%-10s : watcher error' % self.uid)
finally:
self.log.info('%-10s : stop' % self.uid)
def run(self):
try:
# print '%s start' % self.uid
while not self.term.is_set():
# print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 4 and self.pnum == 1:
print "4"
ru.raise_on(self.uid, RAISE_ON)
# print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
except SystemExit:
print '%s exit' % (self.uid)
except KeyboardInterrupt:
print '%s intr' % (self.uid)
finally:
print '%s final' % (self.uid)
def work(self):
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
self.log.info('%-10s : %ds sleep start' % (self.uid, item))
time.sleep(item)
self.log.info('%-10s : %ds sleep stop' % (self.uid, item))
ru.raise_on('work')
def run(self):
try:
self.log.info('%-10s : start' % self.uid)
while not self.term.is_set():
time.sleep(0.1) # start things
ru.raise_on('watch')
for t in self.things:
if not t.is_alive():
self.log.info('%-10s : %s died' % (self.uid, t.uid))
# a child died. We kill the other children and
# terminate.
# self.stop()
return
self.log.info('%-10s : %s ok' % (self.uid, t.uid))
except ThreadExit:
raise RuntimeError('%-10s : watcher exit requested [%s]' % \
(self.uid, self.ident))
except Exception as e:
raise RuntimeError('%-10s : watcher error' % self.uid)
finally:
self.log.info('%-10s : stop' % self.uid)
def __init__(self, name, cfg, verbose):
mt.Thread.__init__(self)
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=self.verbose)
self.cfg = cfg
self.term = mt.Event()
ru.raise_on('init')
def __init__(self, name, cfg, verbose):
mt.Thread.__init__(self)
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=self.verbose)
self.cfg = cfg
self.term = mt.Event()
ru.raise_on('init')
def work(self, units):
if not isinstance(units, list):
units = [units]
self.advance(units, rps.AGENT_EXECUTING, publish=True, push=False)
ru.raise_on('work bulk')
for unit in units:
self._handle_unit(unit)
def work(self, units):
if not isinstance(units, list):
units = [units]
self.advance(units, rps.AGENT_EXECUTING, publish=True, push=False)
ru.raise_on('work bulk')
for unit in units:
self._handle_unit(unit)
def run(self):
# We can't catch SIGINT, for the reasons discussed in the introduction.
# With the default SIGINT handler, SIGINT can hit in unexpected places,
# mostly when thread termination and process termination race. Thus we
# can't use SIGINT at all.
#
# We can, however, use a different signal to communicate termination
# requests from sub-threads to the main thread. Here we use `SIGUSR2`
# (`SIGUSR1` is reserved for debugging purposes in the radical stack).
#
# We also install a `SIGTERM` handler, to initiate orderly shutdown on
# system termination signals.
#
signal.signal(signal.SIGTERM, sigterm_handler)
signal.signal(signal.SIGUSR2, sigusr2_handler)
self.ospid = os.getpid()
self.tid = mt.currentThread().ident
self.uid = "p.%d.0 %8d.%s" % (self.pnum, self.ospid, self.tid)
try:
print '%s start' % self.uid
# create worker thread
self.worker = WorkerThread(self.num, self.pnum, 0)
self.worker.start()
self.watcher = WatcherThread([self.worker], self.num, self.pnum, 1)
self.watcher.start()
while True:
print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 3 and self.pnum == 1:
print "3"
ru.raise_on(self.uid, RAISE_ON)
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
except SystemExit:
print '%s exit' % (self.uid)
except KeyboardInterrupt:
print '%s intr' % (self.uid)
finally:
self.finalize()
def run(self):
# We can't catch SIGINT, for the reasons discussed in the introduction.
# With the default SIGINT handler, SIGINT can hit in unexpected places,
# mostly when thread termination and process termination race. Thus we
# can't use SIGINT at all.
#
# We can, however, use a different signal to communicate termination
# requests from sub-threads to the main thread. Here we use `SIGUSR2`
# (`SIGUSR1` is reserved for debugging purposes in the radical stack).
#
# We also install a `SIGTERM` handler, to initiate orderly shutdown on
# system termination signals.
#
signal.signal(signal.SIGTERM, sigterm_handler)
signal.signal(signal.SIGUSR2, sigusr2_handler)
self.ospid = os.getpid()
self.tid = mt.currentThread().ident
self.uid = "p.%d.0 %8d.%s" % (self.pnum, self.ospid, self.tid)
try:
print '%s start' % self.uid
# create worker thread
self.worker = WorkerThread(self.num, self.pnum, 0)
self.worker.start()
self.watcher = WatcherThread([self.worker], self.num, self.pnum, 1)
self.watcher.start()
while True:
print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 3 and self.pnum == 1:
print "3"
ru.raise_on(self.uid, RAISE_ON)
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
except SystemExit:
print '%s exit' % (self.uid)
except KeyboardInterrupt:
print '%s intr' % (self.uid)
finally:
self.finalize()
def stop(self):
# NOTE: this can be called from the watcher subthread
# make sure the watcher loop is gone
ru.raise_on('stop')
self.term.set() # end watcher loop
ru.raise_on('stop')
# tell children whats up
self._proc_term.set() # end process childs
self._thread_term.set() # end thread childs
ru.raise_on('stop')
for t in self.things:
self.log.info('%-10s : join %s' % (self.uid, t.uid))
t.stop()
t.join(timeout=JOIN_TIMEOUT)
if t.is_alive():
self.log.info('%-10s : kill %s' % (self.uid, t.uid))
## # FIXME: differentiate between procs and threads
## ru.raise_in_thread(tident=t.ident)
## t.join(timeout=JOIN_TIMEOUT)
if t.is_alive():
self.log.info('%-10s : zombied %s' % (self.uid, t.uid))
else:
self.log.info('%-10s : joined %s' % (self.uid, t.uid))
ru.raise_on('stop')
self.log.info('%-10s : stopped' % self.uid)
def stop(self):
# NOTE: this can be called from the watcher subthread
# make sure the watcher loop is gone
ru.raise_on('stop')
self.term.set() # end watcher loop
ru.raise_on('stop')
# tell children whats up
self._proc_term.set() # end process childs
self._thread_term.set() # end thread childs
ru.raise_on('stop')
for t in self.things:
self.log.info('%-10s : join %s' % (self.uid, t.uid))
t.stop()
t.join(timeout=JOIN_TIMEOUT)
if t.is_alive():
self.log.info('%-10s : kill %s' % (self.uid, t.uid))
## # FIXME: differentiate between procs and threads
## ru.raise_in_thread(tident=t.ident)
## t.join(timeout=JOIN_TIMEOUT)
if t.is_alive():
self.log.info('%-10s : zombied %s' % (self.uid, t.uid))
else:
self.log.info('%-10s : joined %s' % (self.uid, t.uid))
ru.raise_on('stop')
self.log.info('%-10s : stopped' % self.uid)
def __init__(self, name, cfg, verbose):
ru.raise_on('init')
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=self.verbose)
self.cfg = cfg
self.things = list()
self.term = None # child only
ru.raise_on('init')
ru.Process.__init__(self, name=self.uid, log=self.log)
def __init__(self, name, cfg, verbose):
ru.raise_on('init')
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=self.verbose)
self.cfg = cfg
self.things = list()
self.term = None # child only
ru.raise_on('init')
ru.Process.__init__(self, name=self.uid, log=self.log)
def main(num):
# *always* install SIGTERM and SIGINT handlers, which will translate those
# signals into exceptable exceptions.
signal.signal(signal.SIGTERM, sigterm_handler)
signal.signal(signal.SIGUSR2, sigusr2_handler)
watcher = None
p1 = None
p2 = None
try:
pid = os.getpid()
tid = mt.currentThread().ident
uid = "m.0.0 %8d.%s" % (pid, tid)
print '%s start' % uid
p1 = ProcessWorker(num, 1)
p2 = ProcessWorker(num, 2)
p1.start()
p2.start()
watcher = WatcherThread([p1, p2], num, 0, 1)
watcher.start()
while True:
print '%s run' % uid
time.sleep(SLEEP)
if num == 1:
print "1"
ru.raise_on(uid, RAISE_ON)
print '%s stop' % uid
except RuntimeError as e:
print '%s error %s [%s]' % (uid, e, type(e))
except SystemExit:
print '%s exit' % (uid)
except KeyboardInterrupt:
print '%s intr' % (uid)
finally:
finalize(p1, p2, watcher)
def run(self):
try:
print '%s start' % self.uid
while not self.term.is_set():
# print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 2 and self.pnum == 0:
print "2"
ru.raise_on(self.uid, RAISE_ON)
if self.num == 5 and self.pnum == 1:
print "5"
ru.raise_on(self.uid, RAISE_ON)
# check watchables
for thing in self.to_watch:
if thing.is_alive():
print '%s event: thing %s is alive' % (self.uid,
thing.uid)
else:
print '%s event: thing %s has died' % (self.uid,
thing.uid)
ru.cancel_main_thread()
assert (False) # we should never get here
# check MainThread
if not self.main.is_alive():
print '%s: main thread gone - terminate' % self.uid
self.stop()
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
ru.cancel_main_thread()
except SystemExit:
print '%s exit' % (self.uid)
# do *not* cancel main thread here! We get here after the cancel
# signal has been sent in the main loop above
finally:
print '%s final' % (self.uid)
def main(num):
# *always* install SIGTERM and SIGINT handlers, which will translate those
# signals into exceptable exceptions.
signal.signal(signal.SIGTERM, sigterm_handler)
signal.signal(signal.SIGUSR2, sigusr2_handler)
watcher = None
p1 = None
p2 = None
try:
pid = os.getpid()
tid = mt.currentThread().ident
uid = "m.0.0 %8d.%s" % (pid, tid)
print '%s start' % uid
p1 = ProcessWorker(num, 1)
p2 = ProcessWorker(num, 2)
p1.start()
p2.start()
watcher = WatcherThread([p1, p2], num, 0, 1)
watcher.start()
while True:
print '%s run' % uid
time.sleep(SLEEP)
if num == 1:
print "1"
ru.raise_on(uid, RAISE_ON)
print '%s stop' % uid
except RuntimeError as e:
print '%s error %s [%s]' % (uid, e, type(e))
except SystemExit:
print '%s exit' % (uid)
except KeyboardInterrupt:
print '%s intr' % (uid)
finally:
finalize(p1, p2, watcher)
def inner_5(arg_1, arg_2): # pylint: disable=W0613
os.environ['RU_RAISE_ON_TEST'] = '3'
for i in range(10):
print(i)
ru.raise_on('test')
print()
os.environ['RU_RAISE_ON_RAND'] = 'RANDOM_10'
for i in range(100):
try:
ru.raise_on('rand')
except Exception:
print('raised on %d' % i)
def run(self):
try:
print '%s start' % self.uid
while not self.term.is_set():
# print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 2 and self.pnum == 0:
print "2"
ru.raise_on(self.uid, RAISE_ON)
if self.num == 5 and self.pnum == 1:
print "5"
ru.raise_on(self.uid, RAISE_ON)
for thing in self.to_watch:
if thing.is_alive():
print '%s event: thing %s is alive' % (self.uid,
thing.uid)
else:
print '%s event: thing %s has died' % (self.uid,
thing.uid)
ru.cancel_main_thread('usr2')
raise RuntimeError('thing %s has died - assert' %
thing.uid)
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
ru.cancel_main_thread('usr2')
except SystemExit:
print '%s exit' % (self.uid)
# do *not* cancel main thread here! We get here
#ru.cancel_main_thread('usr2')
except KeyboardInterrupt:
print '%s intr' % (self.uid)
ru.cancel_main_thread('usr2')
finally:
print '%s final' % (self.uid)
def __init__(self, name, cfg, term, verbose):
mt.Thread.__init__(self)
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=verbose)
self.cfg = cfg
self.term = term
ru.raise_on('init')
# we don't allow subsubthreads
# FIXME: this could be lifted, but we leave in place and
# re-evaluate as needed.
if not ru.is_main_thread():
raise RuntimeError('threads must be spawned by MainThread [%s]' % \
ru.get_thread_name())
def run(self):
try:
print '%s start' % self.uid
while not self.term.is_set():
# print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 2 and self.pnum == 0:
print "2"
ru.raise_on(self.uid, RAISE_ON)
if self.num == 5 and self.pnum == 1:
print "5"
ru.raise_on(self.uid, RAISE_ON)
# check watchables
for thing in self.to_watch:
if thing.is_alive():
print '%s event: thing %s is alive' % (self.uid, thing.uid)
else:
print '%s event: thing %s has died' % (self.uid, thing.uid)
ru.cancel_main_thread()
assert(False) # we should never get here
# check MainThread
if not self.main.is_alive():
print '%s: main thread gone - terminate' % self.uid
self.stop()
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
ru.cancel_main_thread()
except SystemExit:
print '%s exit' % (self.uid)
# do *not* cancel main thread here! We get here after the cancel
# signal has been sent in the main loop above
finally:
print '%s final' % (self.uid)
def __init__(self, name, cfg, term, verbose):
mt.Thread.__init__(self)
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=verbose)
self.cfg = cfg
self.term = term
ru.raise_on('init')
# we don't allow subsubthreads
# FIXME: this could be lifted, but we leave in place and
# re-evaluate as needed.
if not ru.is_main_thread():
raise RuntimeError('threads must be spawned by MainThread [%s]' % \
ru.get_thread_name())
def _handle_unit(self, cu):
ru.raise_on('work unit')
# import pprint
# self._log.info('handle cu: %s', pprint.pformat(cu))
try:
# prep stdout/err so that we can append w/o checking for None
cu['stdout'] = ''
cu['stderr'] = ''
cpt = cu['description']['cpu_process_type']
gpt = cu['description']['gpu_process_type'] # FIXME: use
# FIXME: this switch is insufficient for mixed units (MPI/OpenMP)
if cpt == 'MPI': launcher = self._mpi_launcher
else : launcher = self._task_launcher
if not launcher:
raise RuntimeError("no launcher (process type = %s)" % cpt)
self._log.debug("Launching unit with %s (%s).",
launcher.name, launcher.launch_command)
assert(cu['slots'])
# Start a new subprocess to launch the unit
self.spawn(launcher=launcher, cu=cu)
except Exception as e:
# append the startup error to the units stderr. This is
# not completely correct (as this text is not produced
# by the unit), but it seems the most intuitive way to
# communicate that error to the application/user.
self._log.exception("error running CU")
if cu.get('stderr') is None:
cu['stderr'] = ''
cu['stderr'] += "\nPilot cannot start compute unit:\n%s\n%s" \
% (str(e), traceback.format_exc())
# Free the Slots, Flee the Flots, Ree the Frots!
if cu.get('slots'):
self.publish(rpc.AGENT_UNSCHEDULE_PUBSUB, cu)
self.advance(cu, rps.FAILED, publish=True, push=False)
def _handle_unit(self, cu):
ru.raise_on('work unit')
# import pprint
# self._log.info('handle cu: %s', pprint.pformat(cu))
try:
# prep stdout/err so that we can append w/o checking for None
cu['stdout'] = ''
cu['stderr'] = ''
cpt = cu['description']['cpu_process_type']
gpt = cu['description']['gpu_process_type'] # FIXME: use
# FIXME: this switch is insufficient for mixed units (MPI/OpenMP)
if cpt == 'MPI': launcher = self._mpi_launcher
else: launcher = self._task_launcher
if not launcher:
raise RuntimeError("no launcher (process type = %s)" % cpt)
self._log.debug("Launching unit with %s (%s).", launcher.name,
launcher.launch_command)
assert (cu['slots'])
# Start a new subprocess to launch the unit
self.spawn(launcher=launcher, cu=cu)
except Exception as e:
# append the startup error to the units stderr. This is
# not completely correct (as this text is not produced
# by the unit), but it seems the most intuitive way to
# communicate that error to the application/user.
self._log.exception("error running CU")
if cu.get('stderr') is None:
cu['stderr'] = ''
cu['stderr'] += "\nPilot cannot start compute unit:\n%s\n%s" \
% (str(e), traceback.format_exc())
# Free the Slots, Flee the Flots, Ree the Frots!
if cu.get('slots'):
self.publish(rpc.AGENT_UNSCHEDULE_PUBSUB, cu)
self.advance(cu, rps.FAILED, publish=True, push=False)
def run(self):
try:
print '%s start' % self.uid
while not self.term.is_set():
# print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 2 and self.pnum == 0:
print "2"
ru.raise_on(self.uid, RAISE_ON)
if self.num == 5 and self.pnum == 1:
print "5"
ru.raise_on(self.uid, RAISE_ON)
for thing in self.to_watch:
if thing.is_alive():
print '%s event: thing %s is alive' % (self.uid, thing.uid)
else:
print '%s event: thing %s has died' % (self.uid, thing.uid)
ru.cancel_main_thread('usr2')
raise RuntimeError('thing %s has died - assert' % thing.uid)
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
ru.cancel_main_thread('usr2')
except SystemExit:
print '%s exit' % (self.uid)
# do *not* cancel main thread here! We get here
#ru.cancel_main_thread('usr2')
except KeyboardInterrupt:
print '%s intr' % (self.uid)
ru.cancel_main_thread('usr2')
finally:
print '%s final' % (self.uid)
def inner_5(arg_1, arg_2):
import os
os.environ['RU_RAISE_ON_TEST'] = '3'
for i in range(10):
print i
ru.raise_on('test')
print
import os
os.environ['RU_RAISE_ON_RAND'] = 'RANDOM_10'
for i in range(100):
try:
ru.raise_on('rand')
except Exception as e:
print 'raised on %d' % i
def initialize_child(self):
setproctitle.setproctitle(self.uid)
self.log = ru.get_logger('radical.' + self.uid + '.child',
level=self.verbose)
ru.raise_on('init')
# first create threads and procs to be watched
for name, cfg in self.cfg.iteritems():
self.log.info('child %s: ', name)
if 'child' in name:
child = Child(name=name, cfg=cfg, verbose=self.verbose)
child.start()
self.things.append(child)
elif 'worker' in name:
worker = Worker(name=name, cfg=cfg, verbose=self.verbose)
worker.start()
self.things.append(worker)
ru.raise_on('init')
def stop(self):
ru.raise_on('stop')
assert (self.pid) # child was spanwed
## assert(self.is_parent) # is parent process
## assert(ru.is_main_thread()) # is main thread
self.term.set()
self.log.info('%-10s : stop child' % self.uid)
self.watcher.stop()
ru.raise_on('stop')
## # we check if the watcher finishes.
## if None == ru.watch_condition(cond=self.watcher.is_alive,
## target=False,
## timeout=JOIN_TIMEOUT):
## self.log.info('%-10s : could not stop child - kill' % self.uid)
## self.watcher.kill()
## FIXME: we could attempt a kill and *not* join afterwards, just let py GC
## do the rest
## FIXME: the above is equivalent to `t.join(timeout); t.is_alive()
self.watcher.join(JOIN_TIMEOUT)
self.log.info('%-10s : child stopped (alive: %s)' %
(self.uid, bool(self.is_alive())))
ru.raise_on('stop')
def stop(self):
ru.raise_on('stop')
assert(self.pid) # child was spanwed
## assert(self.is_parent) # is parent process
## assert(ru.is_main_thread()) # is main thread
self.term.set()
self.log.info('%-10s : stop child' % self.uid)
self.watcher.stop()
ru.raise_on('stop')
## # we check if the watcher finishes.
## if None == ru.watch_condition(cond=self.watcher.is_alive,
## target=False,
## timeout=JOIN_TIMEOUT):
## self.log.info('%-10s : could not stop child - kill' % self.uid)
## self.watcher.kill()
## FIXME: we could attempt a kill and *not* join afterwards, just let py GC
## do the rest
## FIXME: the above is equivalent to `t.join(timeout); t.is_alive()
self.watcher.join(JOIN_TIMEOUT)
self.log.info('%-10s : child stopped (alive: %s)' % (self.uid, bool(self.is_alive())))
ru.raise_on('stop')
def work(worker):
# a simple worker routine which sleeps repeatedly for a random number of
# seconds, until a term signal is set. The given 'worker' can be a thread
# or process, or in fact anything which has a self.uid and self.term.
try:
worker.log.info('%-10s : work start' % worker.uid)
while not worker.term.is_set():
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
worker.log.info('%-10s : %ds sleep start' % (worker.uid, item))
time.sleep(item)
worker.log.info('%-10s : %ds sleep stop' % (worker.uid, item))
ru.raise_on('work')
worker.log.info('%-10s : work term requested' % worker.uid)
except Exception as e:
worker.log.info('%-10s : work fail [%s]' % (worker.uid, e))
raise
def work(worker):
# a simple worker routine which sleeps repeatedly for a random number of
# seconds, until a term signal is set. The given 'worker' can be a thread
# or process, or in fact anything which has a self.uid and self.term.
try:
worker.log.info('%-10s : work start' % worker.uid)
while not worker.term.is_set():
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
worker.log.info('%-10s : %ds sleep start' % (worker.uid, item))
time.sleep(item)
worker.log.info('%-10s : %ds sleep stop' % (worker.uid, item))
ru.raise_on('work')
worker.log.info('%-10s : work term requested' % worker.uid)
except Exception as e:
worker.log.info('%-10s : work fail [%s]' % (worker.uid, e))
raise
def _handle_unit(self, cu):
ru.raise_on('work unit')
try:
if cu['description']['mpi']:
launcher = self._mpi_launcher
else:
launcher = self._task_launcher
if not launcher:
raise RuntimeError("no launcher (mpi=%s)" %
cu['description']['mpi'])
self._log.debug("Launching unit with %s (%s).", launcher.name,
launcher.launch_command)
assert (cu['opaque_slots']) # FIXME: no assert, but check
self._prof.prof('exec', msg='unit launch', uid=cu['uid'])
# Start a new subprocess to launch the unit
self.spawn(launcher=launcher, cu=cu)
except Exception as e:
# append the startup error to the units stderr. This is
# not completely correct (as this text is not produced
# by the unit), but it seems the most intuitive way to
# communicate that error to the application/user.
self._log.exception("error running CU")
if cu.get('stderr') is None:
cu['stderr'] = ''
cu['stderr'] += "\nPilot cannot start compute unit:\n%s\n%s" \
% (str(e), traceback.format_exc())
# Free the Slots, Flee the Flots, Ree the Frots!
if cu['opaque_slots']:
self.publish(rpc.AGENT_UNSCHEDULE_PUBSUB, cu)
self.advance(cu, rps.FAILED, publish=True, push=False)
def work(self, units):
if not isinstance(units, list):
units = [units]
self.advance(units, rps.AGENT_STAGING_INPUT, publish=True, push=False)
ru.raise_on('work bulk')
# we first filter out any units which don't need any input staging, and
# advance them again as a bulk. We work over the others one by one, and
# advance them individually, to avoid stalling from slow staging ops.
no_staging_units = list()
staging_units = list()
for unit in units:
# check if we have any staging directives to be enacted in this
# component
actionables = list()
for sd in unit['description'].get('input_staging', []):
if sd['action'] in [rpc.LINK, rpc.COPY, rpc.MOVE, rpc.TARBALL]:
actionables.append(sd)
if actionables:
staging_units.append([unit, actionables])
else:
no_staging_units.append(unit)
if no_staging_units:
self.advance(no_staging_units,
rps.AGENT_SCHEDULING_PENDING,
publish=True,
push=True)
for unit, actionables in staging_units:
self._handle_unit(unit, actionables)
def initialize_child(self):
setproctitle.setproctitle(self.uid)
self.log = ru.get_logger('radical.' + self.uid + '.child',
level=self.verbose)
ru.raise_on('init')
# first create threads and procs to be watched
for name,cfg in self.cfg.iteritems():
self.log.info('child %s: ', name)
if 'child' in name:
child = Child(name=name,
cfg=cfg,
verbose=self.verbose)
child.start()
self.things.append(child)
elif 'worker' in name:
worker = Worker(name=name,
cfg=cfg,
verbose=self.verbose)
worker.start()
self.things.append(worker)
ru.raise_on('init')
def finalize_child(self):
for thing in self.things:
thing.stop()
ru.raise_on('stop')
self.log.info('%-10s : stop child' % self.uid)
ru.raise_on('stop')
self.log.info('%-10s : child stopped' % self.uid)
ru.raise_on('stop')
def finalize_child(self):
for thing in self.things:
thing.stop()
ru.raise_on('stop')
self.log.info('%-10s : stop child' % self.uid)
ru.raise_on('stop')
self.log.info('%-10s : child stopped' % self.uid)
ru.raise_on('stop')
def __init__(self, name, cfg, term, verbose):
ru.raise_on('init')
mp.Process.__init__(self)
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=verbose)
self.is_parent = True
self.cfg = cfg
self.wterm = term # term sig shared with parent watcher
self.term = mp.Event() # private term signal
self.killed = False
# start watcher for own children and threads
ru.raise_on('init')
self.watcher = Watcher(cfg, verbose='error')
self.watcher.start()
ru.raise_on('init')
def __init__(self, name, cfg, term, verbose):
ru.raise_on('init')
mp.Process.__init__(self)
self.uid = name
self.verbose = verbose
self.log = ru.get_logger('radical.' + self.uid, level=verbose)
self.is_parent = True
self.cfg = cfg
self.wterm = term # term sig shared with parent watcher
self.term = mp.Event() # private term signal
self.killed = False
# start watcher for own children and threads
ru.raise_on('init')
self.watcher = Watcher(cfg, verbose='error')
self.watcher.start()
ru.raise_on('init')
def __init__(self, cfg, verbose):
ru.raise_on('init')
mt.Thread.__init__(self)
self.cfg = cfg
self.term = mt.Event()
self._thread_term = mt.Event()
self._proc_term = mp.Event()
self.things = list()
self.uid = None
for name,_ in cfg.iteritems():
if 'watcher' in name:
if self.uid:
raise ValueError('only one watcher supported')
self.uid = name
self.log = ru.get_logger('radical.' + self.uid + '.child',
level=verbose)
ru.raise_on('init')
# first create threads and procs to be watched
for name,_cfg in cfg.iteritems():
self.log.info('child %s: ', name)
if 'child' in name:
child = Child(name=name,
cfg=_cfg,
term=self._proc_term,
verbose=verbose)
child.start()
self.things.append(child)
elif 'worker' in name:
worker = Worker(name=name,
cfg=_cfg,
term=self._thread_term,
verbose=verbose)
worker.start()
self.things.append(worker)
ru.raise_on('init')
def __init__(self, cfg, verbose):
ru.raise_on('init')
mt.Thread.__init__(self)
self.cfg = cfg
self.term = mt.Event()
self._thread_term = mt.Event()
self._proc_term = mp.Event()
self.things = list()
self.uid = None
for name, _ in cfg.iteritems():
if 'watcher' in name:
if self.uid:
raise ValueError('only one watcher supported')
self.uid = name
self.log = ru.get_logger('radical.' + self.uid + '.child',
level=verbose)
ru.raise_on('init')
# first create threads and procs to be watched
for name, _cfg in cfg.iteritems():
self.log.info('child %s: ', name)
if 'child' in name:
child = Child(name=name,
cfg=_cfg,
term=self._proc_term,
verbose=verbose)
child.start()
self.things.append(child)
elif 'worker' in name:
worker = Worker(name=name,
cfg=_cfg,
term=self._thread_term,
verbose=verbose)
worker.start()
self.things.append(worker)
ru.raise_on('init')
def _handle_unit(self, unit, actionables):
ru.raise_on('work unit')
uid = unit['uid']
# NOTE: see documentation of cu['sandbox'] semantics in the ComputeUnit
# class definition.
sandbox = unit['unit_sandbox']
# By definition, this compoentn lives on the pilot's target resource.
# As such, we *know* that all staging ops which would refer to the
# resource now refer to file://localhost, and thus translate the unit,
# pilot and resource sandboxes into that scope. Some assumptions are
# made though:
#
# * paths are directly translatable across schemas
# * resource level storage is in fact accessible via file://
#
# FIXME: this is costly and should be cached.
unit_sandbox = ru.Url(unit['unit_sandbox'])
pilot_sandbox = ru.Url(unit['pilot_sandbox'])
resource_sandbox = ru.Url(unit['resource_sandbox'])
unit_sandbox.schema = 'file'
pilot_sandbox.schema = 'file'
resource_sandbox.schema = 'file'
unit_sandbox.host = 'localhost'
pilot_sandbox.host = 'localhost'
resource_sandbox.host = 'localhost'
src_context = {
'pwd': str(unit_sandbox), # !!!
'unit': str(unit_sandbox),
'pilot': str(pilot_sandbox),
'resource': str(resource_sandbox)
}
tgt_context = {
'pwd': str(unit_sandbox), # !!!
'unit': str(unit_sandbox),
'pilot': str(pilot_sandbox),
'resource': str(resource_sandbox)
}
# we can now handle the actionable staging directives
for sd in actionables:
action = sd['action']
flags = sd['flags']
did = sd['uid']
src = sd['source']
tgt = sd['target']
self._prof.prof('staging_in_start', uid=uid, msg=did)
assert (action in [
rpc.COPY, rpc.LINK, rpc.MOVE, rpc.TRANSFER, rpc.TARBALL
])
# we only handle staging which does *not* include 'client://' src or
# tgt URLs - those are handled by the umgr staging components
if src.startswith('client://') and action != rpc.TARBALL:
self._log.debug('skip staging for src %s', src)
self._prof.prof('staging_in_skip', uid=uid, msg=did)
continue
if tgt.startswith('client://'):
self._log.debug('skip staging for tgt %s', tgt)
self._prof.prof('staging_in_skip', uid=uid, msg=did)
continue
# Fix for when the target PATH is empty
# we assume current directory is the unit staging 'unit://'
# and we assume the file to be copied is the base filename of the source
if tgt is None: tgt = ''
if tgt.strip() == '':
tgt = 'unit:///{}'.format(os.path.basename(src))
# Fix for when the target PATH is exists *and* it is a folder
# we assume the 'current directory' is the target folder
# and we assume the file to be copied is the base filename of the source
elif os.path.exists(tgt.strip()) and os.path.isdir(tgt.strip()):
tgt = os.path.join(tgt, os.path.basename(src))
src = complete_url(src, src_context, self._log)
tgt = complete_url(tgt, tgt_context, self._log)
# Currently, we use the same schema for files and folders.
assert (tgt.schema == 'file'), 'staging tgt must be file://'
if action in [rpc.COPY, rpc.LINK, rpc.MOVE]:
assert (
src.schema == 'file'), 'staging src expected as file://'
# SAGA will take care of dir creation - but we do it manually
# for local ops (copy, link, move)
if flags & rpc.CREATE_PARENTS and action != rpc.TRANSFER:
tgtdir = os.path.dirname(tgt.path)
if tgtdir != sandbox:
self._log.debug("mkdir %s", tgtdir)
rpu.rec_makedir(tgtdir)
if action == rpc.COPY:
try:
shutil.copytree(src.path, tgt.path)
except OSError as exc:
if exc.errno == errno.ENOTDIR:
shutil.copy(src.path, tgt.path)
else:
raise
elif action == rpc.LINK:
# Fix issue/1513 if link source is file and target is folder.
# should support POSIX standard where link is created
# with the same name as the source
if os.path.isfile(src.path) and os.path.isdir(tgt.path):
os.symlink(
src.path,
'%s/%s' % (tgt.path, os.path.basename(src.path)))
else: # default behavior
os.symlink(src.path, tgt.path)
elif action == rpc.MOVE:
shutil.move(src.path, tgt.path)
elif action == rpc.TRANSFER:
# NOTE: TRANSFER directives don't arrive here right now.
# FIXME: we only handle srm staging right now, and only for
# a specific target proxy. Other TRANSFER directives are
# left to umgr input staging. We should use SAGA to
# attempt all staging ops which do not involve the client
# machine.
if src.schema == 'srm':
# FIXME: cache saga handles
srm_dir = rs.filesystem.Directory('srm://proxy/?SFN=bogus')
srm_dir.copy(src, tgt)
srm_dir.close()
else:
self._log.error('no transfer for %s -> %s', src, tgt)
self._prof.prof('staging_in_fail', uid=uid, msg=did)
raise NotImplementedError('unsupported transfer %s' % src)
elif action == rpc.TARBALL:
# If somethig was staged via the tarball method, the tarball is
# extracted and then removed from the unit folder.
self._log.debug('extract tarball for %s', uid)
tar = tarfile.open('%s/%s.tar' %
(os.path.dirname(tgt.path), uid))
tar.extractall(path=os.path.dirname(tgt.path))
tar.close()
# FIXME: make tarball removal dependent on debug settings
# os.remove(os.path.dirname(tgt.path) + '/' + uid + '.tar')
self._prof.prof('staging_in_stop', uid=uid, msg=did)
# all staging is done -- pass on to the scheduler
self.advance(unit,
rps.AGENT_SCHEDULING_PENDING,
publish=True,
push=True)
def run(self):
# We can't catch signals from child processes and threads, so we only
# look out for SIGTERM signals from the parent process. Upon receiving
# such, we'll stop.
#
# We also start a watcher (WatcherThread) which babysits all spawned
# threads and processes, an which will also call stop() on any problems.
# This should then trickle up to the parent, who will also have
# a watcher checking on us.
self.ospid = os.getpid()
self.tid = mt.currentThread().ident
self.uid = "p.%d.0 %8d.%s" % (self.pnum, self.ospid, self.tid)
try:
# ------------------------------------------------------------------
def sigterm_handler(signum, frame):
# on sigterm, we invoke stop(), which will exit.
# Python should (tm) give that signal to the main thread.
# If not, we lost.
assert(mt.currentThread().name == 'MainThread')
self.stop()
# ------------------------------------------------------------------
signal.signal(signal.SIGTERM, sigterm_handler)
print '%s start' % self.uid
# create worker thread
self.worker1 = WorkerThread(self.num, self.pnum, 0)
self.worker1.start()
self.worker2 = WorkerThread(self.num, self.pnum, 0)
self.worker2.start()
self.watcher = WatcherThread([self.worker1, self.worker2],
self.num, self.pnum, 1)
self.watcher.start()
while True:
print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 3 and self.pnum == 1:
print "3"
ru.raise_on(self.uid, RAISE_ON)
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
except SystemExit:
print '%s exit' % (self.uid)
except KeyboardInterrupt:
print '%s intr' % (self.uid)
finally:
# we came here either due to errors in run(), KeyboardInterrupt from
# the WatcherThread, or clean exit. Either way, we stop all
# children.
self.stop()
item = WORK_MIN + (random.random() * (WORK_MAX - WORK_MIN))
self.log.info('%-10s : %ds sleep start' % (self.uid, item))
time.sleep(item)
self.log.info('%-10s : %ds sleep stop' % (self.uid, item))
ru.raise_on('work')
self.log.info('%-10s : work term requested' % self.uid)
except Exception as e:
self.log.info('%-10s : work fail [%s]' % (self.uid, e))
raise
self.log.info('%-10s : thread exit requested' % self.uid)
# ------------------------------------------------------------------------------
#
if __name__ == '__main__':
setproctitle.setproctitle('rp.main')
child = Child(name='root', cfg=config, verbose='debug')
child.start()
ru.raise_on('init')
time.sleep(TIME_ALIVE)
ru.raise_on('stop')
child.stop()
# ------------------------------------------------------------------------------
def _pilot_watcher_cb(self):
# FIXME: we should actually use SAGA job state notifications!
# FIXME: check how race conditions are handles: we may detect
# a finalized SAGA job and change the pilot state -- but that
# pilot may have transitioned into final state via the normal
# notification mechanism already. That probably should be sorted
# out by the pilot manager, which will receive notifications for
# both transitions. As long as the final state is the same,
# there should be no problem anyway. If it differs, the
# 'cleaner' final state should prevail, in this ordering:
# cancel
# timeout
# error
# disappeared
# This implies that we want to communicate 'final_cause'
# we don't want to lock our members all the time. For that reason we
# use a copy of the pilots_tocheck list and iterate over that, and only
# lock other members when they are manipulated.
ru.raise_on('pilot_watcher_cb')
tc = rs.job.Container()
with self._pilots_lock, self._check_lock:
for pid in self._checking:
tc.add(self._pilots[pid]['job'])
states = tc.get_states()
self._log.debug('bulk states: %s', states)
# if none of the states is final, we have nothing to do.
# We can't rely on the ordering of tasks and states in the task
# container, so we hope that the task container's bulk state query lead
# to a caching of state information, and we thus have cache hits when
# querying the pilots individually
final_pilots = list()
with self._pilots_lock, self._check_lock:
for pid in self._checking:
state = self._pilots[pid]['job'].state
self._log.debug('saga job state: %s %s', pid, state)
if state in [rs.job.DONE, rs.job.FAILED, rs.job.CANCELED]:
pilot = self._pilots[pid]['pilot']
if state == rs.job.DONE : pilot['state'] = rps.DONE
if state == rs.job.FAILED : pilot['state'] = rps.FAILED
if state == rs.job.CANCELED: pilot['state'] = rps.CANCELED
final_pilots.append(pilot)
if final_pilots:
for pilot in final_pilots:
with self._check_lock:
# stop monitoring this pilot
self._checking.remove(pilot['uid'])
self._log.debug('final pilot %s %s', pilot['uid'], pilot['state'])
self.advance(final_pilots, push=False, publish=True)
# all checks are done, final pilots are weeded out. Now check if any
# pilot is scheduled for cancellation and is overdue, and kill it
# forcefully.
to_cancel = list()
with self._pilots_lock:
for pid in self._pilots:
pilot = self._pilots[pid]['pilot']
time_cr = pilot.get('cancel_requested')
# check if the pilot is final meanwhile
if pilot['state'] in rps.FINAL:
continue
if time_cr and time_cr + JOB_CANCEL_DELAY < time.time():
self._log.debug('pilot needs killing: %s : %s + %s < %s',
pid, time_cr, JOB_CANCEL_DELAY, time.time())
del(pilot['cancel_requested'])
self._log.debug(' cancel pilot %s', pid)
to_cancel.append(pid)
if to_cancel:
self._kill_pilots(to_cancel)
return True
def work(self, units):
if not isinstance(units, list):
units = [units]
self.advance(units, rps.AGENT_STAGING_OUTPUT, publish=True, push=False)
ru.raise_on('work bulk')
# we first filter out any units which don't need any input staging, and
# advance them again as a bulk. We work over the others one by one, and
# advance them individually, to avoid stalling from slow staging ops.
no_staging_units = list()
staging_units = list()
for unit in units:
uid = unit['uid']
# From here on, any state update will hand control over to the umgr
# again. The next unit update should thus push *all* unit details,
# not only state.
unit['$all'] = True
unit['control'] = 'umgr_pending'
# we always dig for stdout/stderr
self._handle_unit_stdio(unit)
# NOTE: all units get here after execution, even those which did not
# finish successfully. We do that so that we can make
# stdout/stderr available for failed units (see
# _handle_unit_stdio above). But we don't need to perform any
# other staging for those units, and in fact can make them
# final.
if unit['target_state'] != rps.DONE:
unit['state'] = unit['target_state']
self._log.debug('unit %s skips staging (%s)', uid,
unit['state'])
no_staging_units.append(unit)
continue
# check if we have any staging directives to be enacted in this
# component
actionables = list()
for sd in unit['description'].get('output_staging', []):
if sd['action'] in [rpc.LINK, rpc.COPY, rpc.MOVE]:
actionables.append(sd)
if actionables:
# this unit needs some staging
staging_units.append([unit, actionables])
else:
# this unit does not need any staging at this point, and can be
# advanced
unit['state'] = rps.UMGR_STAGING_OUTPUT_PENDING
no_staging_units.append(unit)
if no_staging_units:
self.advance(no_staging_units, publish=True, push=True)
for unit, actionables in staging_units:
self._handle_unit_staging(unit, actionables)
def run(self):
# We can't catch signals from child processes and threads, so we only
# look out for SIGTERM signals from the parent process. Upon receiving
# such, we'll stop.
#
# We also start a watcher (WatcherThread) which babysits all spawned
# threads and processes, an which will also call stop() on any problems.
# This should then trickle up to the parent, who will also have
# a watcher checking on us.
self.ospid = os.getpid()
self.tid = mt.currentThread().ident
self.uid = "p.%d.0 %8d.%s" % (self.pnum, self.ospid, self.tid)
try:
# ------------------------------------------------------------------
def sigterm_handler(signum, frame):
# on sigterm, we invoke stop(), which will exit.
# Python should (tm) give that signal to the main thread.
# If not, we lost.
assert (mt.currentThread().name == 'MainThread')
self.stop()
# ------------------------------------------------------------------
signal.signal(signal.SIGTERM, sigterm_handler)
print '%s start' % self.uid
# create worker thread
self.worker1 = WorkerThread(self.num, self.pnum, 0)
self.worker1.start()
self.worker2 = WorkerThread(self.num, self.pnum, 0)
self.worker2.start()
self.watcher = WatcherThread([self.worker1, self.worker2],
self.num, self.pnum, 1)
self.watcher.start()
while True:
print '%s run' % self.uid
time.sleep(SLEEP)
if self.num == 3 and self.pnum == 1:
print "3"
ru.raise_on(self.uid, RAISE_ON)
print '%s stop' % self.uid
except Exception as e:
print '%s error %s [%s]' % (self.uid, e, type(e))
except SystemExit:
print '%s exit' % (self.uid)
except KeyboardInterrupt:
print '%s intr' % (self.uid)
finally:
# we came here either due to errors in run(), KeyboardInterrupt from
# the WatcherThread, or clean exit. Either way, we stop all
# children.
self.stop()
# self.stop()
return
self.log.info('%-10s : %s ok' % (self.uid, t.uid))
except ThreadExit:
raise RuntimeError('%-10s : watcher exit requested [%s]' % \
(self.uid, self.ident))
except Exception as e:
raise RuntimeError('%-10s : watcher error' % self.uid)
finally:
self.log.info('%-10s : stop' % self.uid)
# ------------------------------------------------------------------------------
#
if __name__ == '__main__':
setproctitle.setproctitle('rp.main')
watcher = Watcher(config, verbose='debug')
watcher.start()
ru.raise_on('init')
time.sleep(TIME_ALIVE)
ru.raise_on('stop')
watcher.stop()
watcher.join()
# ------------------------------------------------------------------------------
def _pilot_watcher_cb(self):
# FIXME: we should actually use SAGA job state notifications!
# FIXME: check how race conditions are handles: we may detect
# a finalized SAGA job and change the pilot state -- but that
# pilot may have transitioned into final state via the normal
# notification mechanism already. That probably should be sorted
# out by the pilot manager, which will receive notifications for
# both transitions. As long as the final state is the same,
# there should be no problem anyway. If it differs, the
# 'cleaner' final state should prevail, in this ordering:
# cancel
# timeout
# error
# disappeared
# This implies that we want to communicate 'final_cause'
# we don't want to lock our members all the time. For that reason we
# use a copy of the pilots_tocheck list and iterate over that, and only
# lock other members when they are manipulated.
ru.raise_on('pilot_watcher_cb')
tc = rs.job.Container()
with self._pilots_lock, self._check_lock:
for pid in self._checking:
tc.add(self._pilots[pid]['job'])
states = tc.get_states()
self._log.debug('bulk states: %s', states)
# if none of the states is final, we have nothing to do.
# We can't rely on the ordering of tasks and states in the task
# container, so we hope that the task container's bulk state query lead
# to a caching of state information, and we thus have cache hits when
# querying the pilots individually
final_pilots = list()
with self._pilots_lock, self._check_lock:
for pid in self._checking:
state = self._pilots[pid]['job'].state
self._log.debug('saga job state: %s %s', pid, state)
if state in [rs.job.DONE, rs.job.FAILED, rs.job.CANCELED]:
pilot = self._pilots[pid]['pilot']
if state == rs.job.DONE: pilot['state'] = rps.DONE
if state == rs.job.FAILED: pilot['state'] = rps.FAILED
if state == rs.job.CANCELED: pilot['state'] = rps.CANCELED
final_pilots.append(pilot)
if final_pilots:
for pilot in final_pilots:
with self._check_lock:
# stop monitoring this pilot
self._checking.remove(pilot['uid'])
self._log.debug('final pilot %s %s', pilot['uid'],
pilot['state'])
self.advance(final_pilots, push=False, publish=True)
# all checks are done, final pilots are weeded out. Now check if any
# pilot is scheduled for cancellation and is overdue, and kill it
# forcefully.
to_cancel = list()
to_advance = list()
with self._pilots_lock:
for pid in self._pilots:
pilot = self._pilots[pid]['pilot']
time_cr = pilot.get('cancel_requested')
# check if the pilot is final meanwhile
if pilot['state'] in rps.FINAL:
continue
if time_cr and time_cr + JOB_CANCEL_DELAY < time.time():
self._log.debug('pilot needs killing: %s : %s + %s < %s',
pid, time_cr, JOB_CANCEL_DELAY,
time.time())
del (pilot['cancel_requested'])
self._log.debug(' cancel pilot %s', pid)
to_cancel.append(pid)
if to_cancel:
self._kill_pilots(to_cancel)
return True
def stop(self):
ru.raise_on('stop')
self.term.set()
ru.raise_on('stop')
def stop(self):
self.term.set()
ru.raise_on('stop')
