def slot_info_to_command(slot_info):
"""
Given a slot_info, creates a command used by gloo to launch a single job.
:param slot_info: host and slot to execute the run command on
:return:
"""
host_name = slot_info.hostname
horovod_rendez_env = ('HOROVOD_HOSTNAME={hostname} '
'HOROVOD_RANK={rank} '
'HOROVOD_SIZE={size} '
'HOROVOD_LOCAL_RANK={local_rank} '
'HOROVOD_LOCAL_SIZE={local_size} '
'HOROVOD_CROSS_RANK={cross_rank} '
'HOROVOD_CROSS_SIZE={cross_size} '.format(
hostname=host_name,
rank=slot_info.rank,
size=slot_info.size,
local_rank=slot_info.local_rank,
local_size=slot_info.local_size,
cross_rank=slot_info.cross_rank,
cross_size=slot_info.cross_size))
return '{horovod_env} {env} {run_command}'.format(
horovod_env=horovod_rendez_env,
env=' '.join([
'%s=%s' % (key, quote(value)) for key, value in env.items()
if env_util.is_exportable(key)
]),
run_command=run_command)
def slot_info_to_command(slot_info):
"""
Given a slot_info, creates a command used by gloo to launch a single job.
:param slot_info: host and slot to execute the run command on
:return:
"""
env_vars = create_slot_env_vars(slot_info)
horovod_rendez_env = " ".join(
[f"{k}={str(v)}" for k, v in env_vars.items()])
return '{shell} {horovod_env} {env} {run_command}' .format(
shell='env',
horovod_env=horovod_rendez_env,
env=' '.join(['%s=%s' % (key, quote(value)) for key, value in env.items()
if env_util.is_exportable(key)]),
run_command=run_command)
def mpi_run(settings, nics, env, command, stdout=None, stderr=None):
"""
Runs mpi_run.
Args:
settings: Settings for running MPI.
Note: settings.num_proc and settings.hosts must not be None.
nics: Interfaces to include by MPI.
env: Environment dictionary to use for running command.
command: Command and arguments to run as a list of string.
stdout: Stdout of the mpi process.
Only used when settings.run_func_mode is True.
stderr: Stderr of the mpi process.
Only used when settings.run_func_mode is True.
"""
if env is not None and not isinstance(env, dict):
raise Exception('env argument must be a dict, not {type}: {env}'
.format(type=type(env), env=env))
mpi_impl_flags, impl_binding_args, mpi = _get_mpi_implementation_flags(settings.tcp_flag, env=env)
if mpi_impl_flags is None:
raise Exception(_MPI_NOT_FOUND_ERROR_MSG)
impi = _IMPI_IMPL == mpi
ssh_args = []
if settings.ssh_port:
ssh_args += [f'-p {settings.ssh_port}']
if settings.ssh_identity_file:
ssh_args += [f'-i {settings.ssh_identity_file}']
mpi_ssh_args = ''
if ssh_args:
joined_ssh_args = ' '.join(ssh_args)
mpi_ssh_args = f'-bootstrap=ssh -bootstrap-exec-args \"{joined_ssh_args}\"' if impi else f'-mca plm_rsh_args \"{joined_ssh_args}\"'
tcp_intf_arg = '-mca btl_tcp_if_include {nics}'.format(
nics=','.join(nics)) if nics and not impi else ''
nccl_socket_intf_arg = '-{opt} NCCL_SOCKET_IFNAME={nics}'.format(
opt='genv' if impi else 'x',
nics=','.join(nics)) if nics else ''
# On large cluster runs (e.g. Summit), we need extra settings to work around OpenMPI issues
host_names, host_to_slots = hosts.parse_hosts_and_slots(settings.hosts)
if not impi and host_names and len(host_names) >= _LARGE_CLUSTER_THRESHOLD:
mpi_impl_flags.append('-mca plm_rsh_no_tree_spawn true')
mpi_impl_flags.append('-mca plm_rsh_num_concurrent {}'.format(len(host_names)))
# if user does not specify any hosts, mpirun by default uses local host.
# There is no need to specify localhost.
hosts_arg = '-{opt} {hosts}'.format(opt='hosts' if impi else 'H',
hosts=','.join(host_names) if host_names and impi else settings.hosts)
ppn_arg = ' '
if host_to_slots and impi:
ppn = host_to_slots[host_names[0]]
for h_name in host_names[1:]:
if ppn != host_to_slots[h_name]:
raise Exception('''Different slots in -hosts parameter are not supported in Intel(R) MPI.
Use -machinefile <machine_file> for this purpose.''')
ppn_arg = ' -ppn {} '.format(ppn)
if settings.prefix_output_with_timestamp and not impi:
mpi_impl_flags.append('--timestamp-output')
binding_args = settings.binding_args if settings.binding_args and not impi else ' '.join(impl_binding_args)
basic_args = '-l' if impi else '--allow-run-as-root --tag-output'
output = []
if settings.output_filename:
output.append('-outfile-pattern' if impi else '--output-filename')
output.append(settings.output_filename)
env_list = '' if impi else ' '.join(
'-x %s' % key for key in sorted(env.keys()) if env_util.is_exportable(key))
# Pass all the env variables to the mpirun command.
mpirun_command = (
'mpirun {basic_args} '
'-np {num_proc}{ppn_arg}{hosts_arg} '
'{binding_args} '
'{mpi_args} '
'{mpi_ssh_args} '
'{tcp_intf_arg} '
'{nccl_socket_intf_arg} '
'{output_filename_arg} '
'{env} {extra_mpi_args} {command}' # expect a lot of environment variables
.format(basic_args=basic_args,
num_proc=settings.num_proc,
ppn_arg=ppn_arg,
hosts_arg=hosts_arg,
binding_args=binding_args,
mpi_args=' '.join(mpi_impl_flags),
tcp_intf_arg=tcp_intf_arg,
nccl_socket_intf_arg=nccl_socket_intf_arg,
mpi_ssh_args=mpi_ssh_args,
output_filename_arg=' '.join(output),
env=env_list,
extra_mpi_args=settings.extra_mpi_args if settings.extra_mpi_args else '',
command=' '.join(quote(par) for par in command))
)
if settings.verbose >= 2:
print(mpirun_command)
# we need the driver's PATH and PYTHONPATH in env to run mpirun,
# env for mpirun is different to env encoded in mpirun_command
for var in ['PATH', 'PYTHONPATH']:
if var not in env and var in os.environ:
# copy env so we do not leak env modifications
env = copy.copy(env)
# copy var over from os.environ
env[var] = os.environ[var]
# Execute the mpirun command.
if settings.run_func_mode:
exit_code = safe_shell_exec.execute(mpirun_command, env=env, stdout=stdout, stderr=stderr)
if exit_code != 0:
raise RuntimeError("mpirun failed with exit code {exit_code}".format(exit_code=exit_code))
else:
os.execve('/bin/sh', ['/bin/sh', '-c', mpirun_command], env)
def mpi_run(settings, nics, env, command, stdout=None, stderr=None):
"""
Runs mpi_run.
Args:
settings: Settings for running MPI.
Note: settings.num_proc and settings.hosts must not be None.
nics: Interfaces to include by MPI.
env: Environment dictionary to use for running command.
command: Command and arguments to run as a list of string.
stdout: Stdout of the mpi process.
Only used when settings.run_func_mode is True.
stderr: Stderr of the mpi process.
Only used when settings.run_func_mode is True.
"""
if env is not None and not isinstance(env, dict):
raise Exception(
'env argument must be a dict, not {type}: {env}'.format(
type=type(env), env=env))
mpi_impl_flags, impl_binding_args = _get_mpi_implementation_flags(
settings.tcp_flag, env=env)
if mpi_impl_flags is None:
raise Exception(_MPI_NOT_FOUND_ERROR_MSG)
ssh_port_arg = '-mca plm_rsh_args \"-p {ssh_port}\"'.format(
ssh_port=settings.ssh_port) if settings.ssh_port else ''
# if user does not specify any hosts, mpirun by default uses local host.
# There is no need to specify localhost.
hosts_arg = '-H {hosts}'.format(hosts=settings.hosts)
tcp_intf_arg = '-mca btl_tcp_if_include {nics}'.format(
nics=','.join(nics)) if nics else ''
nccl_socket_intf_arg = '-x NCCL_SOCKET_IFNAME={nics}'.format(
nics=','.join(nics)) if nics else ''
# On large cluster runs (e.g. Summit), we need extra settings to work around OpenMPI issues
host_names, _ = hosts.parse_hosts_and_slots(settings.hosts)
if host_names and len(host_names) >= _LARGE_CLUSTER_THRESHOLD:
mpi_impl_flags.append('-mca plm_rsh_no_tree_spawn true')
mpi_impl_flags.append('-mca plm_rsh_num_concurrent {}'.format(
len(host_names)))
binding_args = settings.binding_args if settings.binding_args else ' '.join(
impl_binding_args)
# Pass all the env variables to the mpirun command.
mpirun_command = (
'mpirun --allow-run-as-root --tag-output '
'-np {num_proc} {hosts_arg} '
'{binding_args} '
'{mpi_args} '
'{ssh_port_arg} '
'{tcp_intf_arg} '
'{nccl_socket_intf_arg} '
'{output_filename_arg} '
'{env} {extra_mpi_args} {command}' # expect a lot of environment variables
.format(num_proc=settings.num_proc,
hosts_arg=hosts_arg,
binding_args=binding_args,
mpi_args=' '.join(mpi_impl_flags),
tcp_intf_arg=tcp_intf_arg,
nccl_socket_intf_arg=nccl_socket_intf_arg,
ssh_port_arg=ssh_port_arg,
output_filename_arg='--output-filename ' +
settings.output_filename if settings.output_filename else '',
env=' '.join('-x %s' % key for key in sorted(env.keys())
if env_util.is_exportable(key)),
extra_mpi_args=settings.extra_mpi_args
if settings.extra_mpi_args else '',
command=' '.join(quote(par) for par in command)))
if settings.verbose >= 2:
print(mpirun_command)
# we need the driver's PATH and PYTHONPATH in env to run mpirun,
# env for mpirun is different to env encoded in mpirun_command
for var in ['PATH', 'PYTHONPATH']:
if var not in env and var in os.environ:
# copy env so we do not leak env modifications
env = copy.copy(env)
# copy var over from os.environ
env[var] = os.environ[var]
# Execute the mpirun command.
if settings.run_func_mode:
exit_code = safe_shell_exec.execute(mpirun_command,
env=env,
stdout=stdout,
stderr=stderr)
if exit_code != 0:
raise RuntimeError(
"mpirun failed with exit code {exit_code}".format(
exit_code=exit_code))
else:
os.execve('/bin/sh', ['/bin/sh', '-c', mpirun_command], env)