def test_numa_multi_instance_run_command(
mock_system, mock_subprocess, mock_platform, mock_os, mock_open,
mock_path_exists, test_num_instances, test_socket_id, test_cpu_list, expected_cpu_bind):
""" Test the multi instance run using numactl by trying different combinations of
cpu lists and the number of cores used per instance. Checks the system call that
is run to verify that it matches the cpu groups that are expected. """
platform_util = MagicMock(cpu_core_list=test_cpu_list)
test_output_dir = "/tmp/output"
args = MagicMock(verbose=True, model_name=test_model_name, batch_size=100,
numa_cores_per_instance=test_num_instances, precision="fp32",
output_dir=test_output_dir, socket_id=test_socket_id)
os.environ["PYTHON_EXE"] = "python"
os.environ["MPI_HOSTNAMES"] = "None"
os.environ["MPI_NUM_PROCESSES"] = "None"
base_model_init = BaseModelInitializer(args, [], platform_util)
mock_path_exists.return_value = True
expected_omp_num_threads = "OMP_NUM_THREADS={}".format(test_num_instances)
# call run_command and then check the output
test_run_command = "python foo.py"
base_model_init.run_command(test_run_command)
system_call_args = mock_system.call_args[0][0]
assert expected_omp_num_threads in system_call_args
for cpu_bind in expected_cpu_bind:
assert "numactl --localalloc --physcpubind={} {} >> {}".\
format(cpu_bind, test_run_command, test_output_dir) in system_call_args
def test_command_prefix_tcmalloc_fp32(precision, mock_glob):
""" FP32 models should have TC Malloc disabled by default, but models should
include LD_PRELOAD in the command prefix if disable_tcmalloc=False is explicitly set. """
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
test_tcmalloc_lib = "/usr/lib/libtcmalloc.so.4.2.6"
mock_glob.return_value = [test_tcmalloc_lib]
os.environ["PYTHON_EXE"] = "python"
args.socket_id = 0
args.precision = precision
# By default, TCMalloc should not be used
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id)
assert "LD_PRELOAD={}".format(test_tcmalloc_lib) not in command_prefix
assert "numactl --cpunodebind=0 --membind=0" in command_prefix
# If tcmalloc is disabled, LD_PRELOAD shouild not be in the prefix
args.disable_tcmalloc = False
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id)
assert "LD_PRELOAD={}".format(test_tcmalloc_lib) in command_prefix
assert "numactl --cpunodebind=0 --membind=0" in command_prefix
# If numactl is set to false, we should not have numactl in the prefix
args.disable_tcmalloc = True
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id,
numactl=False)
assert "LD_PRELOAD={}".format(test_tcmalloc_lib) not in command_prefix
assert "numactl" not in command_prefix
def test_command_prefix_tcmalloc_int8(precision, mock_glob):
""" For Int8 models, TCMalloc should be enabled by default and models should include
LD_PRELOAD in the command prefix, unless disable_tcmalloc=True is set """
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
test_tcmalloc_lib = "/usr/lib/libtcmalloc.so.4.2.6"
mock_glob.return_value = [test_tcmalloc_lib]
os.environ["PYTHON_EXE"] = "python"
os.environ["MPI_NUM_PROCESSES"] = "None"
args.socket_id = 0
args.precision = precision
# If tcmalloc is not disabled, we should have LD_PRELOAD in the prefix
args.disable_tcmalloc = False
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id)
assert "LD_PRELOAD={}".format(test_tcmalloc_lib) in command_prefix
assert "numactl --cpunodebind=0 --membind=0" in command_prefix
# If tcmalloc is disabled, LD_PRELOAD shouild not be in the prefix
args.disable_tcmalloc = True
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id)
assert "LD_PRELOAD={}".format(test_tcmalloc_lib) not in command_prefix
assert "numactl --cpunodebind=0 --membind=0" in command_prefix
# If numactl is set to false, we should not have numactl in the prefix
args.disable_tcmalloc = False
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id,
numactl=False)
assert "LD_PRELOAD={}".format(test_tcmalloc_lib) in command_prefix
assert "numactl" not in command_prefix
def test_set_kmp_vars_config_json_exists(mock_json):
"""Test config.json when exists"""
# Setup base model init with test settings
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
os.environ["PYTHON_EXE"] = "python"
base_model_init = BaseModelInitializer(args, [], platform_util)
file_descriptor, config_file_path = tempfile.mkstemp(suffix=".json")
base_model_init.set_kmp_vars(config_file_path)
def test_base_model_initializer(mock_system, mock_subprocess, mock_platform,
mock_os):
# Setup base model init with test settings
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
os.environ["PYTHON_EXE"] = "python"
base_model_init = BaseModelInitializer(args, [], platform_util)
# call run_command and then check the output
test_run_command = "python foo.py"
base_model_init.run_command(test_run_command)
mock_system.assert_called_with(test_run_command)
def test_numa_multi_instance_run_command(mock_system, mock_subprocess,
mock_platform, mock_os, mock_open,
mock_glob, mock_path_exists,
test_num_instances, test_socket_id,
test_cpu_list, expected_cpu_bind,
precision):
""" Test the multi instance run using numactl by trying different combinations of
cpu lists and the number of cores used per instance. Checks the system call that
is run to verify that it matches the cpu groups that are expected. """
platform_util = MagicMock(cpu_core_list=test_cpu_list)
test_output_dir = "/tmp/output"
args = MagicMock(verbose=True,
model_name=test_model_name,
batch_size=100,
numa_cores_per_instance=test_num_instances,
precision=precision,
output_dir=test_output_dir,
mode="inference",
socket_id=test_socket_id,
benchmark_only=True)
os.environ["PYTHON_EXE"] = "python"
os.environ["MPI_HOSTNAMES"] = "None"
os.environ["MPI_NUM_PROCESSES"] = "None"
ld_preload_path = "/usr/lib/libtcmalloc.so.4.2.6"
mock_glob.return_value = [ld_preload_path]
base_model_init = BaseModelInitializer(args, [], platform_util)
mock_path_exists.return_value = True
expected_omp_num_threads = "OMP_NUM_THREADS={}".format(test_num_instances)
# tcmalloc is used by default for int8, so the LD_PRELOAD path should be set
expected_ld_preload = ""
if precision == "int8":
expected_ld_preload = "LD_PRELOAD={} ".format(ld_preload_path)
test_run_command = "python foo.py"
# For int8, add on the LD_PRELOAD prefix. The run command should know how to move this prefix to the
# front of the command (before numactl)
if precision == "int8":
test_run_command_with_prefix = expected_ld_preload + test_run_command
else:
test_run_command_with_prefix = test_run_command
# call run_command and then check the output
base_model_init.run_command(test_run_command_with_prefix)
system_call_args = mock_system.call_args[0][0]
for cpu_bind in expected_cpu_bind:
expected_cmd = "{}{} numactl --localalloc --physcpubind={} {} >> {}".\
format(expected_ld_preload, expected_omp_num_threads, cpu_bind, test_run_command, test_output_dir)
assert expected_cmd in system_call_args
def test_base_model_initializer(
mock_system, mock_subprocess, mock_platform, mock_os, mock_open):
# Setup base model init with test settings
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name, batch_size=100,
numa_cores_per_instance=None)
os.environ["PYTHON_EXE"] = "python"
os.environ["MPI_HOSTNAMES"] = "None"
os.environ["MPI_NUM_PROCESSES"] = "None"
base_model_init = BaseModelInitializer(args, [], platform_util)
# call run_command and then check the output
test_run_command = "python foo.py"
base_model_init.run_command(test_run_command)
mock_system.assert_called_with(test_run_command)
def test_base_model_initializer(mock_system, mock_subprocess, mock_platform,
mock_os):
setup_mock_values(mock_platform, mock_os, mock_subprocess)
launch_benchmark = LaunchBenchmark()
args, _ = launch_benchmark.parse_args(example_req_args)
test_run_command = "python foo.py"
platform_util = MagicMock()
# Setup base model init with test settings
args.verbose = True
args.model_name = test_model_name
base_model_init = BaseModelInitializer(args, [], platform_util)
# call run_command and then check the output
base_model_init.run_command(test_run_command)
mock_system.assert_called_with(test_run_command)
def test_set_kmp_vars_config_json_does_not_exists():
"""Test config.json does not exist"""
# Setup base model init with test settings
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
os.environ["PYTHON_EXE"] = "python"
base_model_init = BaseModelInitializer(args, [], platform_util)
config_file_path = '/test/foo/config.json'
with catch_stdout() as caught_output:
base_model_init.set_kmp_vars(config_file_path)
output = caught_output.getvalue()
assert "Warning: File {} does not exist and \
cannot be used to set KMP environment variables".format(config_file_path) == output.strip()
def test_command_prefix_numa_multi_instance(precision, mock_glob):
""" Tests that models don't get a numactl command prefix from
base_model_init.get_command_prefix, if numa_cores_per_instance is set, since numactl
commands will be added later for each instance run. """
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
test_tcmalloc_lib = "/usr/lib/libtcmalloc.so.4.2.6"
mock_glob.return_value = [test_tcmalloc_lib]
os.environ["PYTHON_EXE"] = "python"
args.socket_id = 0
args.precision = precision
args.numa_cores_per_instance = 4
base_model_init = BaseModelInitializer(args, [], platform_util)
command_prefix = base_model_init.get_command_prefix(args.socket_id)
assert "numactl" not in command_prefix
def test_num_inter_intra_threads_settings(
mock_system, mock_subprocess, mock_platform, mock_os, mock_open,
mock_path_exists, test_num_instances, test_socket_id, test_num_cores,
test_cpu_list, test_cpuset, expected_inter_threads,
expected_intra_threads):
"""
Tests the base model init function that determines the num_inter_threads and
num_intra_thread values.
"""
platform_util = MagicMock(cpu_core_list=test_cpu_list,
cpuset_cpus=test_cpuset,
num_cores_per_socket=len(test_cpu_list[0]))
test_output_dir = "/tmp/output"
args = MagicMock(verbose=True,
model_name=test_model_name,
batch_size=100,
numa_cores_per_instance=test_num_instances,
precision="fp32",
output_dir=test_output_dir,
socket_id=test_socket_id,
num_cores=test_num_cores,
num_inter_threads=None,
num_intra_threads=None)
os.environ["PYTHON_EXE"] = "python"
os.environ["MPI_HOSTNAMES"] = "None"
os.environ["MPI_NUM_PROCESSES"] = "None"
base_model_init = BaseModelInitializer(args, [], platform_util)
mock_path_exists.return_value = True
# Get the number of inter/intra threads and compared to the expected values
base_model_init.set_num_inter_intra_threads()
print(base_model_init.args.num_inter_threads)
print(base_model_init.args.num_intra_threads)
assert base_model_init.args.num_inter_threads == expected_inter_threads
assert base_model_init.args.num_intra_threads == expected_intra_threads
def test_multi_instance_train_prefix():
platform_util = MagicMock()
args = MagicMock(verbose=True, model_name=test_model_name)
args.num_processes = 2
args.num_processes_per_node = 1
base_model_init = BaseModelInitializer(args, [], platform_util)
command = base_model_init.get_multi_instance_train_prefix(option_list=["--genv:test"])
assert command == "mpirun -n 2 -ppn 1 --genv test "
args.num_processes = None
args.num_processes_per_node = None
base_model_init = BaseModelInitializer(args, [], platform_util)
command = base_model_init.get_multi_instance_train_prefix(option_list=["--genv:test", "--genv:test2"])
assert command == "mpirun --genv test --genv test2 "
