def gloo_run_elastic(settings, env, command):
# Make the output directory if it does not exist
if settings.output_filename:
_mkdir_p(settings.output_filename)
rendezvous = RendezvousServer(settings.verbose)
driver = ElasticDriver(rendezvous,
settings.discovery,
settings.min_np,
settings.max_np,
timeout=settings.elastic_timeout,
verbose=settings.verbose)
handler = create_rendezvous_handler(driver)
global_rendezv_port = rendezvous.start_server(handler)
# Host-to-host common interface detection requires at least 2 hosts in an elastic job.
min_hosts = _get_min_start_hosts(settings)
current_hosts = driver.wait_for_available_slots(settings.num_proc,
min_hosts=min_hosts)
nics = driver_service.get_common_interfaces(
settings, current_hosts.host_assignment_order)
server_ip = network.get_driver_ip(nics)
exec_command = _exec_command_fn(settings)
event = register_shutdown_event()
run_command = get_run_command(command,
server_ip,
nics,
global_rendezv_port,
elastic=True)
create_worker = _create_elastic_worker_fn(exec_command, run_command, env,
event)
driver.start(settings.num_proc, create_worker)
res = driver.get_results()
driver.stop()
rendezvous.stop_server()
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 launch_gloo_elastic(command, exec_command, settings, env,
get_common_interfaces, rendezvous):
# Make the output directory if it does not exist
if settings.output_filename:
_mkdir_p(settings.output_filename)
driver = ElasticDriver(rendezvous,
settings.discovery,
settings.min_np,
settings.max_np,
timeout=settings.elastic_timeout,
reset_limit=settings.reset_limit,
verbose=settings.verbose)
handler = create_rendezvous_handler(driver)
global_rendezv_port = rendezvous.start(handler)
driver.wait_for_available_slots(settings.num_proc)
nics = get_common_interfaces(driver)
server_ip = network.get_driver_ip(nics)
event = register_shutdown_event()
run_command = get_run_command(command,
server_ip,
nics,
global_rendezv_port,
elastic=True)
create_worker = _create_elastic_worker_fn(exec_command, run_command, env,
event)
driver.start(settings.num_proc, create_worker)
res = driver.get_results()
driver.stop()
rendezvous.stop()
if res.error_message is not None:
raise RuntimeError(res.error_message)
for name, value in sorted(res.worker_results.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_server(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()
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_server()