def troubleshoot_server(self):
if self.settings['ssh']:
if servers[self.server]['cluster_soft'].lower() == 'oge':
# delete present server run
logging.error(
'Job {name} has server status {stat} on {server}. Troubleshooting by changing node.'
.format(name=self.job_name,
stat=self.job_status[0],
server=self.server))
ssh = SSH_Client(self.server)
ssh.send_command_to_server(command=delete_command[servers[
self.server]['cluster_soft']] + ' ' + str(self.job_id))
# find available nodes
stdout, _ = ssh.send_command_to_server(
command=list_available_nodes_command[servers[self.server]
['cluster_soft']])
for line in stdout:
node = line.split()[0].split('.')[0].split('node')[1]
if servers[self.server][
'cluster_soft'] == 'OGE' and '0/0/8' in line and node not in self.server_nodes:
self.server_nodes.append(node)
break
else:
logging.error(
'Could not find an available node on the server')
# TODO: continue troubleshooting; if all else fails, put job to sleep for x min and try again searching for a node
return
# modify submit file
content = ssh.read_remote_file(
remote_path=self.remote_path,
filename=submit_filename[servers[self.server]
['cluster_soft']])
for i, line in enumerate(content):
if '#$ -l h=node' in line:
content[i] = '#$ -l h=node{0}.cluster'.format(node)
break
else:
content.insert(7, '#$ -l h=node{0}.cluster'.format(node))
content = ''.join(
content
) # convert list into a single string, not to upset paramico
# resubmit
ssh.upload_file(remote_file_path=os.path.join(
self.remote_path,
submit_filename[servers[self.server]['cluster_soft']]),
file_string=content)
self.run()
elif servers[self.server]['cluster_soft'].lower() == 'slurm':
# TODO: change node on Slurm
# delete present server run
logging.error(
'Job {name} has server status {stat} on {server}. Re-running job.'
.format(name=self.job_name,
stat=self.job_status[0],
server=self.server))
ssh = SSH_Client(self.server)
ssh.send_command_to_server(command=delete_command[servers[
self.server]['cluster_soft']] + ' ' + str(self.job_id))
# resubmit
self.run()
def _upload_submit_file(self):
ssh = SSH_Client(self.server)
ssh.send_command_to_server(
command='mkdir -p {0}'.format(self.remote_path))
remote_file_path = os.path.join(
self.remote_path,
submit_filename[servers[self.server]['cluster_soft']])
ssh.upload_file(remote_file_path=remote_file_path,
file_string=self.submit)
def _upload_input_file(self):
ssh = SSH_Client(self.server)
ssh.send_command_to_server(
command='mkdir -p {0}'.format(self.remote_path))
remote_file_path = os.path.join(self.remote_path,
input_filename[self.software])
ssh.upload_file(remote_file_path=remote_file_path,
file_string=self.input)
self.initial_time = ssh.get_last_modified_time(
remote_file_path=remote_file_path)
def _get_additional_job_info(self):
"""
Download the additional information of stdout and stderr from the server
"""
lines1, lines2 = list(), list()
content = ''
ssh = SSH_Client(self.server)
cluster_soft = servers[self.server]['cluster_soft'].lower()
if cluster_soft in ['oge', 'sge']:
remote_file_path = os.path.join(self.remote_path, 'out.txt')
local_file_path1 = os.path.join(self.local_path, 'out.txt')
try:
ssh.download_file(remote_file_path=remote_file_path, local_file_path=local_file_path1)
except (TypeError, IOError) as e:
logging.warning('Got the following error when trying to download out.txt for {0}:'.format(self.job_name))
logging.warning(e.message)
remote_file_path = os.path.join(self.remote_path, 'err.txt')
local_file_path2 = os.path.join(self.local_path, 'err.txt')
try:
ssh.download_file(remote_file_path=remote_file_path, local_file_path=local_file_path2)
except (TypeError, IOError) as e:
logging.warning('Got the following error when trying to download err.txt for {0}:'.format(self.job_name))
logging.warning(e.message)
if os.path.isfile(local_file_path1):
with open(local_file_path1, 'r') as f:
lines1 = f.readlines()
if os.path.isfile(local_file_path2):
with open(local_file_path2, 'r') as f:
lines2 = f.readlines()
content += ''.join([line for line in lines1])
content += '\n'
content += ''.join([line for line in lines2])
elif cluster_soft == 'slurm':
respond = ssh.send_command_to_server(command='ls -alF', remote_path=self.remote_path)
files = list()
for line in respond[0][0].splitlines():
files.append(line.split()[-1])
for file in files:
if 'slurm' in file and '.out' in file:
remote_file_path = os.path.join(self.remote_path, file)
local_file_path = os.path.join(self.local_path, file)
try:
ssh.download_file(remote_file_path=remote_file_path, local_file_path=local_file_path)
except (TypeError, IOError) as e:
logging.warning('Got the following error when trying to download {0} for {1}:'.format(
file, self.job_name))
logging.warning(e.message)
if os.path.isfile(local_file_path):
with open(local_file_path, 'r') as f:
lines1 = f.readlines()
content += ''.join([line for line in lines1])
content += '\n'
return content