def launch_gloo(command, exec_command, settings, nics, env, server_ip):
"""
Launches the given command multiple times using gloo.
Each command is launched via exec_command.
:param command: command to launch
:param exec_command: means to execute a single command
:param settings: settings for the distribution
:param nics: common interfaces
:param env: environment to use
:param server_ip: ip to use for rendezvous server
"""
# Make the output directory if it does not exist
if settings.output_filename:
_mkdir_p(settings.output_filename)
# start global rendezvous server and get port that it is listening on
rendezvous = RendezvousServer(settings.verbose)
# allocate processes into slots
hosts = parse_hosts(settings.hosts)
host_alloc_plan = get_host_assignments(hosts, settings.num_proc)
# start global rendezvous server and get port that it is listening on
global_rendezv_port = rendezvous.start()
rendezvous.init(host_alloc_plan)
run_command = get_run_command(command, server_ip, nics,
global_rendezv_port)
slot_info_to_command = _slot_info_to_command_fn(run_command, env)
event = register_shutdown_event()
args_list = [[slot_info_to_command(slot_info), slot_info, [event]]
for slot_info in host_alloc_plan]
# If an error occurs in one thread, entire process will be terminated.
# Otherwise, threads will keep running.
res = threads.execute_function_multithreaded(exec_command,
args_list,
block_until_all_done=True)
for name, value in sorted(res.items(), key=lambda item: item[1][1]):
exit_code, timestamp = value
if exit_code != 0:
raise RuntimeError(
'Horovod detected that one or more processes exited with non-zero '
'status, thus causing the job to be terminated. The first process '
'to do so was:\nProcess name: {name}\nExit code: {code}\n'.
format(name=name, code=exit_code))
def test_worker_notification_manager(self):
"""Tests that host add events are sent to the worker notification service and consumed."""
slots = {'host-1': 2}
discovery = FixedHosts(slots)
rendezvous = RendezvousServer()
driver = ElasticDriver(rendezvous, discovery, min_np=2, max_np=4)
driver.wait_for_available_slots(min_np=2)
handler = create_rendezvous_handler(driver)
common_intfs = network.get_local_intfs()
addr = network.get_driver_ip(common_intfs)
port = rendezvous.start(handler)
nic = list(common_intfs)[0]
rank_results = {}
class NotificationReceiver:
def __init__(self):
self.events = []
def on_hosts_updated(self, timestamp):
self.events.append(timestamp)
def add_host():
slots = {'host-1': 2, 'host-2': 2}
discovery.set(slots)
def remove_host():
slots = {'host-2': 2}
discovery.set(slots)
def exec_command(slot_info, events):
manager = WorkerNotificationManager()
manager.init(rendezvous_addr=addr,
rendezvous_port=port,
nic=nic,
hostname=slot_info.hostname,
local_rank=slot_info.local_rank)
notification_receiver = NotificationReceiver()
manager.register_listener(notification_receiver)
driver.record_ready(slot_info.hostname, slot_info.local_rank)
if slot_info.rank == 0:
add_host()
driver.wait_for_available_slots(4)
if slot_info.rank == 0:
remove_host()
# Busy wait for the number of available slots to decrease
while driver._host_manager.current_hosts.count_available_slots(
) > 2:
time.sleep(0.01)
rank_results[slot_info.rank] = notification_receiver.events
return 0, time.time()
driver.start(np=2, create_worker_fn=exec_command)
res = driver.get_results().worker_results
driver.stop()
assert len(res) == 2
for name, (exit_code, timestamp) in res.items():
assert exit_code == 0, name
assert len(rank_results) == 2
for rank, timestamps in rank_results.items():
expected = 2 if rank == 0 else 0
assert len(timestamps) == expected, rank
rendezvous.stop()