def cleanup(tb_hdfs_path):
hopshdfs.log('Performing cleanup')
handle = hopshdfs.get()
if not tb_hdfs_path == None and not tb_hdfs_path == '' and handle.exists(
tb_hdfs_path):
handle.delete(tb_hdfs_path)
hopshdfs.kill_logger()
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_pid = 0
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
try:
#Arguments
if args_dict:
argcount = six.get_function_code(map_fun).co_argcount
names = six.get_function_code(map_fun).co_varnames
args = []
argIndex = 0
param_string = ''
while argcount > 0:
#Get args for executor and run function
param_name = names[argIndex]
param_val = args_dict[param_name][executor_num]
param_string += str(param_name) + '=' + str(
param_val) + '.'
args.append(param_val)
argcount -= 1
argIndex += 1
param_string = param_string[:-1]
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, param_string, 'launcher')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print(
'-------------------------------------------------------')
print('Started running task ' + param_string + '\n')
hopshdfs.log('Started running task ' + param_string)
task_start = datetime.datetime.now()
map_fun(*args)
task_end = datetime.datetime.now()
time_str = 'Finished task ' + param_string + ' - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print(
'-------------------------------------------------------')
hopshdfs.log(time_str)
else:
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'launcher')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print(
'-------------------------------------------------------')
print('Started running task\n')
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = map_fun()
task_end = datetime.datetime.now()
if retval:
_handle_return(retval, hdfs_exec_logdir)
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print(
'-------------------------------------------------------')
hopshdfs.log(time_str)
except:
#Always do cleanup
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
try:
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
except:
pass
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
task_index = None
try:
host = util._get_ip_address()
tmp_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tmp_socket.bind(('', 0))
port = tmp_socket.getsockname()[1]
client = allreduce_reservation.Client(server_addr)
host_port = host + ":" + str(port)
client.register({"worker": host_port, "index": executor_num})
cluster = client.await_reservations()
tmp_socket.close()
client.close()
task_index = _find_index(host_port, cluster)
cluster["task"] = {"type": "worker", "index": task_index}
os.environ["TF_CONFIG"] = json.dumps(cluster)
if task_index == 0:
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'collective_all_reduce')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
if task_index == 0:
hopshdfs.log(gpu_str)
print(gpu_str)
print('-------------------------------------------------------')
print('Started running task \n')
if task_index == 0:
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = map_fun()
if task_index == 0:
if retval:
_handle_return(retval, hdfs_exec_logdir)
task_end = datetime.datetime.now()
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print('-------------------------------------------------------')
if task_index == 0:
hopshdfs.log(time_str)
except:
#Always do cleanup
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
if task_index == 0:
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _wrapper_fun(iter):
for i in iter:
executor_num = i
tb_pid = 0
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices.print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
global local_logdir_bool
try:
#Arguments
if args_dict:
argcount = six.get_function_code(map_fun).co_argcount
names = six.get_function_code(map_fun).co_varnames
args = []
argIndex = 0
param_string = ''
while argcount > 0:
#Get args for executor and run function
param_name = names[argIndex]
param_val = args_dict[param_name][executor_num]
param_string += str(param_name) + '=' + str(
param_val) + '.'
args.append(param_val)
argcount -= 1
argIndex += 1
param_string = param_string[:-1]
val = _get_metric(param_string, app_id, generation_id, run_id)
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs.create_directories(
app_id,
run_id,
param_string,
'differential_evolution',
sub_type='generation.' + str(generation_id))
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs.init_logger()
tb_hdfs_path, tb_pid = tensorboard.register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir_bool)
gpu_str = '\nChecking for GPUs in the environment' + devices.get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print(
'-------------------------------------------------------')
print('Started running task ' + param_string + '\n')
if val:
print('Reading returned metric from previous run: ' +
str(val))
hopshdfs.log('Started running task ' + param_string)
task_start = datetime.datetime.now()
if not val:
val = map_fun(*args)
task_end = datetime.datetime.now()
time_str = 'Finished task ' + param_string + ' - took ' + util.time_diff(
task_start, task_end)
print('\n' + time_str)
hopshdfs.log(time_str)
try:
castval = int(val)
except:
raise ValueError(
'Your function needs to return a metric (number) which should be maximized or minimized'
)
metric_file = hdfs_exec_logdir + '/metric'
fs_handle = hopshdfs.get_fs()
try:
fd = fs_handle.open_file(metric_file, mode='w')
except:
fd = fs_handle.open_file(metric_file, flags='w')
fd.write(str(float(val)).encode())
fd.flush()
fd.close()
print('Returning metric ' + str(val))
print(
'-------------------------------------------------------')
except:
#Always do cleanup
if tb_hdfs_path:
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
if local_logdir_bool:
local_tb = tensorboard.local_logdir_path
util.store_local_tensorboard(local_tb, hdfs_exec_logdir)
hopshdfs.log('Finished running')
if tb_hdfs_path:
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
role = None
client = parameter_server_reservation.Client(server_addr)
try:
host = util._get_ip_address()
tmp_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tmp_socket.bind(('', 0))
port = tmp_socket.getsockname()[1]
host_port = host + ":" + str(port)
exec_spec = {}
if executor_num < num_ps:
exec_spec["task_type"] = "ps"
else:
exec_spec["task_type"] = "worker"
exec_spec["host_port"] = host_port
exec_spec["gpus_present"] = devices.get_num_gpus() > 0
client.register(exec_spec)
cluster = client.await_reservations()
tmp_socket.close()
role, index = _find_task_and_index(host_port, cluster)
cluster_spec = {}
cluster_spec["cluster"] = cluster
cluster_spec["task"] = {"type": role, "index": index}
print(cluster_spec)
os.environ["TF_CONFIG"] = json.dumps(cluster_spec)
if role == "chief":
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'parameter_server')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
if role == "chief":
hopshdfs.log(gpu_str)
print(gpu_str)
print('-------------------------------------------------------')
print('Started running task \n')
if role == "chief":
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = None
if role == "ps":
ps_thread = threading.Thread(target=lambda: map_fun())
ps_thread.start()
print("waiting for workers")
client.await_all_workers_finished()
print("waiting finished")
else:
retval = map_fun()
if role == "chief":
if retval:
_handle_return(retval, hdfs_exec_logdir)
task_end = datetime.datetime.now()
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print('-------------------------------------------------------')
if role == "chief":
hopshdfs.log(time_str)
except:
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
if role == "chief":
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
try:
if role == "worker" or role == "chief":
client.register_worker_finished()
client.close()
except:
pass
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _wrapper_fun(iter):
for i in iter:
executor_num = i
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs.create_directories(
app_id, run_id, None, 'horovod')
tb_pid = 0
tb_hdfs_path = ''
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs.init_logger()
hopshdfs.log('Starting Spark executor with arguments')
if executor_num == 0:
tb_hdfs_path, tb_pid = tensorboard.register(
hdfs_exec_logdir,
hdfs_appid_logdir,
0,
local_logdir=local_logdir)
gpu_str = '\n\nChecking for GPUs in the environment\n' + devices.get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
#1. Download notebook file
fs_handle = hopshdfs.get_fs()
try:
fd = fs_handle.open_file(nb_path, flags='r')
except:
fd = fs_handle.open_file(nb_path, mode='r')
notebook = ''
for line in fd:
notebook += line
path, filename = os.path.split(nb_path)
f_nb = open(filename, "w+")
f_nb.write(notebook)
f_nb.flush()
f_nb.close()
# 2. Convert notebook to py file
jupyter_runnable = os.path.abspath(
os.path.join(os.environ['PYSPARK_PYTHON'], os.pardir)) + '/jupyter'
conversion_cmd = jupyter_runnable + ' nbconvert --to python ' + filename
conversion = subprocess.Popen(conversion_cmd,
shell=True,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE)
conversion.wait()
stdout, stderr = conversion.communicate()
print(stdout)
print(stderr)
# 3. Make py file runnable
py_runnable = os.getcwd() + '/' + filename.split('.')[0] + '.py'
st = os.stat(py_runnable)
os.chmod(py_runnable, st.st_mode | stat.S_IEXEC)
t_gpus = threading.Thread(
target=devices.print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t_gpus.start()
mpi_logfile_path = os.getcwd() + '/mpirun.log'
if os.path.exists(mpi_logfile_path):
os.remove(mpi_logfile_path)
mpi_logfile = open(mpi_logfile_path, 'w')
# 4. Run allreduce
mpi_np = os.environ['MPI_NP']
mpi_cmd = 'HOROVOD_TIMELINE=' + tensorboard.logdir() + '/timeline.json' + \
' TENSORBOARD_LOGDIR=' + tensorboard.logdir() + \
' mpirun -np ' + str(mpi_np) + \
' -bind-to none -map-by slot ' + \
' -x HOROVOD_TIMELINE ' + \
' -x TENSORBOARD_LOGDIR ' + \
' -x NCCL_DEBUG=INFO ' + \
os.environ['PYSPARK_PYTHON'] + ' ' + py_runnable
mpi = subprocess.Popen(mpi_cmd,
shell=True,
stdout=mpi_logfile,
stderr=mpi_logfile,
preexec_fn=util.on_executor_exit('SIGTERM'))
t_log = threading.Thread(target=print_log)
t_log.start()
mpi.wait()
if devices.get_num_gpus() > 0:
t_gpus.do_run = False
t_gpus.join()
return_code = mpi.returncode
if local_logdir:
local_tb = tensorboard.local_logdir_path
pydoop.hdfs.put(local_tb, hdfs_exec_logdir)
if return_code != 0:
cleanup(tb_hdfs_path)
t_log.do_run = False
t_log.join()
raise Exception('mpirun FAILED, look in the logs for the error')
cleanup(tb_hdfs_path)
t_log.do_run = False
t_log.join()
hopshdfs.kill_logger()
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_pid = 0
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
try:
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'mirrored')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print('-------------------------------------------------------')
print('Started running task\n')
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = map_fun()
task_end = datetime.datetime.now()
if retval:
_handle_return(retval, hdfs_exec_logdir)
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print('-------------------------------------------------------')
hopshdfs.log(time_str)
except:
#Always do cleanup
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
try:
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
except:
pass
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def log(message):
hdfs.log(str(message))
