def exec(self):
"""Run the SuperBench benchmarks locally."""
for benchmark_name in self._sb_benchmarks:
if benchmark_name not in self._sb_enabled:
continue
benchmark_config = self._sb_benchmarks[benchmark_name]
benchmark_results = list()
self.__create_benchmark_dir(benchmark_name)
cwd = os.getcwd()
os.chdir(self.__get_benchmark_dir(benchmark_name))
monitor = None
if self.__get_rank_id(
) == 0 and self._sb_monitor_config and self._sb_monitor_config.enable:
if self.__get_platform() == Platform.CUDA:
monitor = Monitor(
None, int(self._sb_monitor_config.sample_duration
or 10),
int(self._sb_monitor_config.sample_interval or 1),
self.__get_monitor_path(benchmark_name))
monitor.start()
else:
logger.warning(
'Monitor can not support ROCM/CPU platform.')
benchmark_real_name = benchmark_name.split(':')[0]
for framework in benchmark_config.frameworks or [
Framework.NONE.value
]:
if benchmark_real_name == 'model-benchmarks' or (
':' not in benchmark_name
and benchmark_name.endswith('_models')):
for model in benchmark_config.models:
full_name = f'{benchmark_name}/{framework}-{model}'
logger.info('Executor is going to execute %s.',
full_name)
context = BenchmarkRegistry.create_benchmark_context(
model,
platform=self.__get_platform(),
framework=Framework(framework.lower()),
parameters=self.__get_arguments(
benchmark_config.parameters))
result = self.__exec_benchmark(full_name, context)
benchmark_results.append(result)
else:
full_name = benchmark_name
logger.info('Executor is going to execute %s.', full_name)
context = BenchmarkRegistry.create_benchmark_context(
benchmark_real_name,
platform=self.__get_platform(),
framework=Framework(framework.lower()),
parameters=self.__get_arguments(
benchmark_config.parameters))
result = self.__exec_benchmark(full_name, context)
benchmark_results.append(result)
if monitor:
monitor.stop()
self.__write_benchmark_results(benchmark_name, benchmark_results)
os.chdir(cwd)
def test_kernel_launch_overhead():
"""Test kernel-launch benchmark."""
context = BenchmarkRegistry.create_benchmark_context(
'kernel-launch',
parameters='--num_warmup 200 --num_steps 20000 --interval 100')
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (benchmark.name == 'kernel-launch')
assert (benchmark.type == BenchmarkType.MICRO)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.num_warmup == 200)
assert (benchmark._args.num_steps == 20000)
assert (benchmark._args.interval == 100)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
assert ('raw_output_0' in benchmark.raw_data)
assert (len(benchmark.raw_data['raw_output_0']) == 1)
assert (isinstance(benchmark.raw_data['raw_output_0'][0], str))
for metric in ['event_time', 'wall_time']:
assert (metric in benchmark.result)
assert (len(benchmark.result[metric]) == 1)
assert (isinstance(benchmark.result[metric][0], numbers.Number))
def test_pytorch_computation_communication_overlap_normal():
"""Test pytorch-computation-communication-overlap benchmark on distributed normal case."""
context = BenchmarkRegistry.create_benchmark_context(
'computation-communication-overlap',
parameters='--num_warmup 5 --num_steps 10 --ratio 5',
framework=Framework.PYTORCH
)
world_size = 2
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
results = utils.simulated_ddp_distributed_benchmark(context, world_size)
assert (results)
for benchmark in results:
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, ComputationCommunicationOverlap))
assert (benchmark.name == 'pytorch-computation-communication-overlap')
assert (benchmark.type == BenchmarkType.MICRO)
# Check predefined parameters of sharding-matmul benchmark.
assert (benchmark._args.kernel == [ComputationKernelType.MUL, ComputationKernelType.MATMUL])
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.num_steps == 10)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
assert (len(benchmark.raw_data) == len(benchmark._args.kernel))
assert (len(benchmark.result) == len(benchmark._args.kernel) + benchmark.default_metric_count)
def test_pytorch_computation_communication_overlap_fake_distributed():
"""Test pytorch-computation-communication-overlap benchmark on single gpu."""
context = BenchmarkRegistry.create_benchmark_context(
'computation-communication-overlap',
parameters='--num_warmup 5 --num_steps 10 --ratio 5',
framework=Framework.PYTORCH
)
port = network.get_free_port()
assert (port)
utils.setup_simulated_ddp_distributed_env(1, 0, port)
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, ComputationCommunicationOverlap))
assert (benchmark.name == 'pytorch-computation-communication-overlap')
assert (benchmark.type == BenchmarkType.MICRO)
# Check predefined parameters of sharding-matmul benchmark.
assert (benchmark._args.kernel == [ComputationKernelType.MUL, ComputationKernelType.MATMUL])
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.num_steps == 10)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
assert (len(benchmark.raw_data) == len(benchmark._args.kernel))
assert (len(benchmark.result) == len(benchmark._args.kernel) + benchmark.default_metric_count)
utils.clean_simulated_ddp_distributed_env()
def test_pytorch_matmul():
"""Test pytorch-matmul benchmark."""
context = BenchmarkRegistry.create_benchmark_context(
'matmul',
platform=Platform.CUDA,
parameters='--run_count 2 --num_steps 20',
framework=Framework.PYTORCH)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (benchmark.name == 'pytorch-matmul')
assert (benchmark.type == BenchmarkType.MICRO)
# Check predefined parameters of sharding-matmul benchmark.
assert (benchmark._args.mode == [ShardingMode.NOSHARDING])
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.run_count == 2)
assert (benchmark._args.num_steps == 20)
# Check results and metrics.
assert (benchmark.run_count == 2)
assert (benchmark.return_code == ReturnCode.SUCCESS)
assert (len(benchmark.raw_data['nosharding_time']) == benchmark.run_count)
assert (len(
benchmark.raw_data['nosharding_time'][0]) == benchmark._args.num_steps)
assert (len(benchmark.result['nosharding_time']) == benchmark.run_count)
def benchmark_list_params_command_handler(name=None):
"""List parameters for benchmarks which match the regular expression.
Args:
name (str, optional): Benchmark name or regular expression. Defaults to None.
Raises:
CLIError: If cannot find the matching benchmark.
"""
for benchmark_name in benchmark_list_command_handler(name):
format_help = ''
for platform in Platform:
if platform in BenchmarkRegistry.benchmarks[benchmark_name]:
format_help = BenchmarkRegistry.get_benchmark_configurable_settings(
BenchmarkRegistry.create_benchmark_context(benchmark_name, platform=platform)
)
break
print(
(
f'=== {benchmark_name} ===\n\n'
f'{format_help}\n\n'
f'default values:\n'
f'{pformat(BenchmarkRegistry.benchmarks[benchmark_name]["predefine_param"])}\n'
)
)
def test_register_benchmark():
"""Test interface BenchmarkRegistry.register_benchmark()."""
# Register the benchmark for all platform if use default platform.
BenchmarkRegistry.register_benchmark('accumulation', AccumulationBenchmark)
for platform in Platform:
context = BenchmarkRegistry.create_benchmark_context('accumulation',
platform=platform)
assert (BenchmarkRegistry.is_benchmark_registered(context))
# Register the benchmark for CUDA platform if use platform=Platform.CUDA.
BenchmarkRegistry.register_benchmark('accumulation-cuda',
AccumulationBenchmark,
platform=Platform.CUDA)
context = BenchmarkRegistry.create_benchmark_context(
'accumulation-cuda', platform=Platform.CUDA)
assert (BenchmarkRegistry.is_benchmark_registered(context))
context = BenchmarkRegistry.create_benchmark_context(
'accumulation-cuda', platform=Platform.ROCM)
assert (BenchmarkRegistry.is_benchmark_registered(context) is False)
def test_is_benchmark_context_valid():
"""Test interface BenchmarkRegistry.is_benchmark_context_valid()."""
# Positive case.
context = BenchmarkRegistry.create_benchmark_context('accumulation',
platform=Platform.CPU)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
# Negative case.
context = 'context'
assert (BenchmarkRegistry.is_benchmark_context_valid(context) is False)
context = None
assert (BenchmarkRegistry.is_benchmark_context_valid(context) is False)
def test_pytorch_bert_base():
"""Test pytorch-bert-base benchmark."""
context = BenchmarkRegistry.create_benchmark_context(
'bert-base',
platform=Platform.CUDA,
parameters=
'--batch_size 1 --num_classes 5 --seq_len 8 --num_warmup 2 --num_steps 4 \
--model_action train inference',
framework=Framework.PYTORCH)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, PytorchBERT))
assert (benchmark.name == 'pytorch-bert-base')
assert (benchmark.type == BenchmarkType.MODEL)
# Check predefined parameters of resnet101 model.
assert (benchmark._args.hidden_size == 768)
assert (benchmark._args.num_hidden_layers == 12)
assert (benchmark._args.num_attention_heads == 12)
assert (benchmark._args.intermediate_size == 3072)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.batch_size == 1)
assert (benchmark._args.num_classes == 5)
assert (benchmark._args.seq_len == 8)
assert (benchmark._args.num_warmup == 2)
assert (benchmark._args.num_steps == 4)
# Check dataset scale.
assert (len(benchmark._dataset) == benchmark._args.sample_count *
benchmark._world_size)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
for metric in [
'fp32_train_step_time', 'fp32_train_throughput',
'fp16_train_step_time', 'fp16_train_throughput',
'fp32_inference_step_time', 'fp32_inference_throughput',
'fp16_inference_step_time', 'fp16_inference_throughput'
]:
assert (len(benchmark.raw_data[metric]) == benchmark.run_count)
assert (len(
benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
assert (len(benchmark.result[metric]) == benchmark.run_count)
def test_tcp_connectivity(self):
"""Test tcp-connectivity benchmark."""
context = BenchmarkRegistry.create_benchmark_context(
'tcp-connectivity',
parameters=
'--hostfile /tmp/superbench/hostfile.test --port 80 --parallel 2',
)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, TCPConnectivityBenchmark))
assert (benchmark.name == 'tcp-connectivity')
assert (benchmark.type == BenchmarkType.MICRO)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.hostfile == '/tmp/superbench/hostfile.test')
assert (benchmark._args.port == 80)
assert (benchmark._args.count == 10)
assert (benchmark._args.timeout == 1)
assert (benchmark._args.parallel == 2)
print(benchmark.result)
assert (benchmark.result)
# Check results and metrics.
assert (benchmark.result['api.github.com_successed_count'][0] == 10)
assert (benchmark.result['api.github.com_failed_count'][0] == 0)
assert (benchmark.result['api.github.com_success_rate'][0] == 100.0)
assert (isinstance(benchmark.result['api.github.com_time_min'][0],
numbers.Number))
assert (isinstance(benchmark.result['api.github.com_time_max'][0],
numbers.Number))
assert (isinstance(benchmark.result['api.github.com_time_avg'][0],
numbers.Number))
assert (isinstance(benchmark.result['localhost_successed_count'][0],
numbers.Number))
assert (isinstance(benchmark.result['localhost_failed_count'][0],
numbers.Number))
assert (isinstance(benchmark.result['localhost_time_max'][0],
numbers.Number))
assert (isinstance(benchmark.result['localhost_time_min'][0],
numbers.Number))
assert (isinstance(benchmark.result['localhost_time_avg'][0],
numbers.Number))
assert (benchmark.return_code == ReturnCode.SUCCESS)
def run_pytorch_lstm(parameters='', check_metrics=[]):
"""Test pytorch-lstm benchmark."""
context = BenchmarkRegistry.create_benchmark_context(
'lstm',
platform=Platform.CUDA,
parameters=parameters,
framework=Framework.PYTORCH)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, PytorchLSTM))
assert (benchmark.name == 'pytorch-lstm')
assert (benchmark.type == BenchmarkType.MODEL)
# Check predefined parameters of lstm model.
assert (benchmark._args.input_size == 256)
assert (benchmark._args.hidden_size == 1024)
assert (benchmark._args.num_layers == 8)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.batch_size == 1)
assert (benchmark._args.num_classes == 5)
assert (benchmark._args.seq_len == 8)
assert (benchmark._args.num_warmup == 2)
assert (benchmark._args.num_steps == 4)
# Check dataset scale.
assert (len(benchmark._dataset) == benchmark._args.sample_count *
benchmark._world_size)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
for metric in check_metrics:
assert (len(benchmark.raw_data[metric]) == benchmark.run_count)
assert (len(
benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
assert (len(benchmark.result[metric]) == benchmark.run_count)
def test_get_benchmark_configurable_settings():
"""Test BenchmarkRegistry interface.
BenchmarkRegistry.get_benchmark_configurable_settings().
"""
# Register benchmarks for testing.
BenchmarkRegistry.register_benchmark('accumulation', AccumulationBenchmark)
context = BenchmarkRegistry.create_benchmark_context('accumulation',
platform=Platform.CPU)
settings = BenchmarkRegistry.get_benchmark_configurable_settings(context)
expected = """optional arguments:
--duration int The elapsed time of benchmark in seconds.
--log_raw_data Log raw data into file instead of saving it into result
object.
--lower_bound int The lower bound for accumulation.
--run_count int The run count of benchmark.
--upper_bound int The upper bound for accumulation."""
assert (settings == expected)
def run_pytorch_cnn(models=[], parameters='', check_metrics=[]):
"""Run pytorch cnn benchmarks."""
for model in models:
context = BenchmarkRegistry.create_benchmark_context(
model,
platform=Platform.CUDA,
parameters=parameters,
framework=Framework.PYTORCH)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, PytorchCNN))
assert (benchmark.name == 'pytorch-' + model)
assert (benchmark.type == BenchmarkType.MODEL)
# Check predefined parameters of resnet101 model.
assert (benchmark._args.model_type == model)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.batch_size == 1)
assert (benchmark._args.image_size == 224)
assert (benchmark._args.num_classes == 5)
assert (benchmark._args.num_warmup == 2)
assert (benchmark._args.num_steps == 4)
# Check Dataset.
assert (len(benchmark._dataset) == benchmark._args.sample_count *
benchmark._world_size)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
for metric in check_metrics:
assert (len(benchmark.raw_data[metric]) == benchmark.run_count)
assert (len(
benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
assert (len(benchmark.result[metric]) == benchmark.run_count)
def create_benchmark(params='--num_steps 8'):
"""Register and create benchmark."""
# Register the FakeModelBenchmark benchmark.
BenchmarkRegistry.register_benchmark(
'pytorch-fake-model',
FakeModelBenchmark,
parameters='--hidden_size 2',
platform=Platform.CUDA,
)
context = BenchmarkRegistry.create_benchmark_context(
'fake-model',
platform=Platform.CUDA,
parameters=params,
framework=Framework.PYTORCH)
name = BenchmarkRegistry._BenchmarkRegistry__get_benchmark_name(context)
assert (name)
(benchmark_class, predefine_params
) = BenchmarkRegistry._BenchmarkRegistry__select_benchmark(
name, context.platform)
assert (benchmark_class)
return benchmark_class(name, predefine_params + ' ' + context.parameters)
def test_pytorch_sharding_matmul():
"""Test pytorch-sharding-matmul benchmark."""
context = BenchmarkRegistry.create_benchmark_context(
'sharding-matmul',
platform=Platform.CUDA,
parameters='--run_count 2 --num_steps 20',
framework=Framework.PYTORCH)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
port = network.get_free_port()
assert (port)
utils.setup_simulated_ddp_distributed_env(1, 0, port)
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (isinstance(benchmark, ShardingMatmul))
assert (benchmark.name == 'pytorch-sharding-matmul')
assert (benchmark.type == BenchmarkType.MICRO)
# Check predefined parameters of sharding-matmul benchmark.
assert (benchmark._args.mode == [
ShardingMode.ALLREDUCE, ShardingMode.ALLGATHER
])
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.run_count == 2)
assert (benchmark._args.num_steps == 20)
# Check results and metrics.
assert (benchmark.run_count == 2)
assert (benchmark.return_code == ReturnCode.SUCCESS)
for metric in ['allreduce_time', 'allgather_time']:
assert (len(benchmark.raw_data[metric]) == benchmark.run_count)
assert (len(
benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
assert (len(benchmark.result[metric]) == benchmark.run_count)
utils.clean_simulated_ddp_distributed_env()
def test_get_benchmark_name():
"""Test interface BenchmarkRegistry.get_benchmark_name()."""
# Register benchmarks for testing.
benchmark_names = [
'accumulation', 'pytorch-accumulation', 'tf1-accumulation',
'onnxruntime-accumulation'
]
for name in benchmark_names:
BenchmarkRegistry.register_benchmark(name, AccumulationBenchmark)
# Test benchmark name for different Frameworks.
benchmark_frameworks = [
Framework.NONE, Framework.PYTORCH, Framework.TENSORFLOW1,
Framework.ONNXRUNTIME
]
for i in range(len(benchmark_names)):
context = BenchmarkRegistry.create_benchmark_context(
'accumulation',
platform=Platform.CPU,
framework=benchmark_frameworks[i])
name = BenchmarkRegistry._BenchmarkRegistry__get_benchmark_name(
context)
assert (name == benchmark_names[i])
def test_pytorch_empty_cache():
"""Test PytorchBase class."""
# Register mnist benchmark.
BenchmarkRegistry.register_benchmark('pytorch-mnist', PytorchMNIST)
# Test cache empty by manually calling torch.cuda.empty_cache().
parameters = '--batch_size 32 --num_warmup 8 --num_steps 64 --model_action train'
benchmark = PytorchMNIST('pytorch-mnist', parameters=parameters)
assert (benchmark)
assert (benchmark._preprocess())
assert (benchmark._benchmark())
del benchmark
assert (torch.cuda.memory_stats()['reserved_bytes.all.current'] > 0)
torch.cuda.empty_cache()
assert (torch.cuda.memory_stats()['reserved_bytes.all.current'] == 0)
# Test automatic cache empty.
context = BenchmarkRegistry.create_benchmark_context(
'pytorch-mnist', parameters='--batch_size 32 --num_warmup 8 --num_steps 64 --model_action train'
)
benchmark = BenchmarkRegistry.launch_benchmark(context)
assert (benchmark)
assert (torch.cuda.memory_stats()['reserved_bytes.all.current'] == 0)
import argparse
from superbench.benchmarks import Platform, Framework, BenchmarkRegistry
from superbench.common.utils import logger
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--distributed',
action='store_true',
default=False,
help='Whether to enable distributed training.')
args = parser.parse_args()
# Specify the model name and benchmark parameters.
model_name = 'gpt2-large'
parameters = '--batch_size 1 --duration 120 --seq_len 128 --precision float32 --run_count 2'
if args.distributed:
parameters += ' --distributed_impl ddp --distributed_backend nccl'
# Create context for gpt2-large benchmark and run it for 120 * 2 seconds.
context = BenchmarkRegistry.create_benchmark_context(
model_name,
platform=Platform.CUDA,
parameters=parameters,
framework=Framework.PYTORCH)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
def test_cublas_functions():
"""Test cublas-function benchmark."""
# Test for default configuration
context = BenchmarkRegistry.create_benchmark_context(
'cublas-function',
platform=Platform.CUDA,
parameters='--num_warmup 10 --num_steps 10 --num_in_step 100')
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (benchmark.name == 'cublas-function')
assert (benchmark.type == BenchmarkType.MICRO)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.num_warmup == 10)
assert (benchmark._args.num_steps == 10)
assert (benchmark._args.num_in_step == 100)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
assert ('raw_output_0' in benchmark.raw_data)
assert (len(benchmark.raw_data['raw_output_0']) == 1)
assert (isinstance(benchmark.raw_data['raw_output_0'][0], str))
assert (19 <= len(benchmark.result))
for metric in list(benchmark.result.keys()):
assert (len(benchmark.result[metric]) == 1)
assert (isinstance(benchmark.result[metric][0], numbers.Number))
if metric != 'return_code':
assert (len(
benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
# Test for custom configuration
custom_config_str = '{"name":"cublasCgemm","m":512,"n":512,"k":32,"transa":1,"transb":0}'
context = BenchmarkRegistry.create_benchmark_context(
'cublas-function',
platform=Platform.CUDA,
parameters=
'--num_warmup 10 --num_steps 10 --num_in_step 100 --config_json_str ' +
custom_config_str)
assert (BenchmarkRegistry.is_benchmark_context_valid(context))
benchmark = BenchmarkRegistry.launch_benchmark(context)
# Check basic information.
assert (benchmark)
assert (benchmark.name == 'cublas-function')
assert (benchmark.type == BenchmarkType.MICRO)
# Check parameters specified in BenchmarkContext.
assert (benchmark._args.num_warmup == 10)
assert (benchmark._args.num_steps == 10)
assert (benchmark._args.num_in_step == 100)
# Check results and metrics.
assert (benchmark.run_count == 1)
assert (benchmark.return_code == ReturnCode.SUCCESS)
assert ('raw_output_0' in benchmark.raw_data)
assert (len(benchmark.raw_data['raw_output_0']) == 1)
assert (isinstance(benchmark.raw_data['raw_output_0'][0], str))
assert (1 + benchmark.default_metric_count == len(benchmark.result))
for metric in list(benchmark.result.keys()):
assert (len(benchmark.result[metric]) == 1)
assert (isinstance(benchmark.result[metric][0], numbers.Number))
if metric != 'return_code':
assert (len(
benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for disk performance.
Commands to run:
python3 examples/benchmarks/memory_bw_latency_performance.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'cpu-memory-bw-latency',
platform=Platform.CPU,
parameters='--tests bandwidth_matrix latency_matrix max_bandwidth')
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for GPU copy bandwidth performance.
Commands to run:
python3 examples/benchmarks/gpu_copy_bw_performance.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'gpu-copy-bw',
platform=Platform.CUDA,
parameters='--mem_type htod dtoh dtod --copy_type sm dma')
# For ROCm environment, please specify the benchmark name and the platform as the following.
# context = BenchmarkRegistry.create_benchmark_context(
# 'gpu-copy-bw', platform=Platform.ROCM, parameters='--mem_type htod dtoh dtod --copy_type sm dma'
# )
# For bidirectional test, please specify parameters as the following.
# parameters='--mem_type htod dtod --copy_type sm dma --bidirectional'
# To enable data checking, please add '--check_data'.
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Microbenchmark benchmark example for TCP connectivity.
Commands to run:
python3 examples/benchmarks/tcp_connectivity.py
"""
from superbench.benchmarks import BenchmarkRegistry
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'tcp-connectivity', parameters='--hostfile /tmp/superbench/hostfile.test --port 80 --parallel 1'
)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info(
'benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result
)
)
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for cudnn performance benchmark.
Commands to run:
python3 examples/benchmarks/cudnn_function.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
parameters = '--num_warmup 8 --num_steps 100 --num_in_step 1000'
context = BenchmarkRegistry.create_benchmark_context(
'cudnn-function', platform=Platform.CUDA, parameters=parameters)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for GPU-Burn.
Commands to run:
python3 examples/benchmarks/gpu_burn_test.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'gpu-burn', platform=Platform.CUDA, parameters='--doubles --tensor_core --time 10'
)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info(
'benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result
)
)
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for sharding-matmul with pytorch.
Commands to run:
python3 -m torch.distributed.launch --nproc_per_node=8 examples/benchmarks/sharding_matmul.py
"""
from superbench.benchmarks import Framework, BenchmarkRegistry
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'sharding-matmul', parameters='--num_steps 20', framework=Framework.PYTORCH
)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info(
'benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result
)
)
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Model benchmark example for Cutlass GEMM FLOPs performance.
Commands to run:
python3 examples/benchmarks/gemm_flops_cuda_performance.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
parameters = '--n 16384 --k 16384 --m 16384'
context = BenchmarkRegistry.create_benchmark_context(
'gemm-flops', platform=Platform.CUDA, parameters=parameters)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Model benchmark example for kernel launch overhead.
Commands to run:
python3 examples/benchmarks/kernel_launch_overhead.py
"""
from superbench.benchmarks import BenchmarkRegistry
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context('kernel-launch')
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for ONNXRuntime inference performance.
Commands to run:
python3 examples/benchmarks/ort_inference_performance.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'ort-inference',
platform=Platform.CUDA,
parameters='--pytorch_models resnet50 resnet101 --precision float16')
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for IB loopback performance.
Commands to run:
python examples/benchmarks/ib_loopback_performance_performance.py
"""
from superbench.benchmarks import BenchmarkRegistry
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context('ib-loopback')
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""Micro benchmark example for device memory bandwidth performance.
Commands to run:
python3 examples/benchmarks/rocm_memory_bw_performance.py
"""
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger
if __name__ == '__main__':
context = BenchmarkRegistry.create_benchmark_context(
'mem-bw', platform=Platform.ROCM)
benchmark = BenchmarkRegistry.launch_benchmark(context)
if benchmark:
logger.info('benchmark: {}, return code: {}, result: {}'.format(
benchmark.name, benchmark.return_code, benchmark.result))
