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 end(metric=None):
global running
global experiment_json
global elastic_id
global driver_tensorboard_hdfs_path
global app_id
if not running:
raise RuntimeError(
"An experiment is not running. Did you forget to call experiment.end()?"
)
try:
if metric:
experiment_json = util.finalize_experiment(experiment_json, None,
str(metric))
util.put_elastic(hopshdfs.project_name(), app_id, elastic_id,
experiment_json)
else:
experiment_json = util.finalize_experiment(experiment_json, None,
None)
util.put_elastic(hopshdfs.project_name(), app_id, elastic_id,
experiment_json)
except:
exception_handler()
raise
finally:
elastic_id += 1
running = False
handle = hopshdfs.get()
if tensorboard.tb_pid != 0:
subprocess.Popen(["kill", str(tensorboard.tb_pid)])
if tensorboard.local_logdir_bool:
local_tb = tensorboard.local_logdir_path
util.store_local_tensorboard(local_tb, tensorboard.events_logdir)
if not tensorboard.endpoint == None and not tensorboard.endpoint == '' \
and handle.exists(tensorboard.endpoint):
handle.delete(tensorboard.endpoint)
hopshdfs.kill_logger()
def _cleanup(tb_hdfs_path):
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):
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()