def _CreateMetadataDict(benchmark_spec):
"""Create metadata dict to be used in run results.
Args:
benchmark_spec: benchmark spec
Returns:
metadata dict
"""
vm = benchmark_spec.vms[0]
metadata = dict()
if benchmark_spec.device == GPU:
metadata.update(cuda_toolkit_8.GetMetadata(vm))
metadata['num_gpus'] = benchmark_spec.num_gpus
metadata['forward_only'] = benchmark_spec.forward_only
metadata['model'] = benchmark_spec.model
metadata['data_name'] = benchmark_spec.data_name
metadata['batch_size'] = benchmark_spec.batch_size
metadata['variable_update'] = benchmark_spec.variable_update
metadata['local_parameter_device'] = benchmark_spec.local_parameter_device
metadata['device'] = benchmark_spec.device
metadata['data_format'] = benchmark_spec.data_format
metadata['use_nccl'] = benchmark_spec.use_nccl
metadata['distortions'] = benchmark_spec.distortions
return metadata
def Run(benchmark_spec):
"""Runs the Stencil2D benchmark. GPU clock speeds must be set already.
Args:
benchmark_spec: The benchmark specification. Contains all data that is
required to run the benchmark.
Returns:
A list of sample.Sample objects.
"""
vms = benchmark_spec.vms
num_gpus = benchmark_spec.num_gpus
master_vm = vms[0]
num_iterations = FLAGS.stencil2d_iterations
problem_sizes = FLAGS.stencil2d_problem_sizes
num_processes = len(vms) * num_gpus
metadata = {}
metadata.update(cuda_toolkit_8.GetMetadata(master_vm))
metadata['benchmark_version'] = BENCHMARK_VERSION
metadata['num_iterations'] = num_iterations
metadata['num_nodes'] = len(vms)
metadata['num_processes'] = num_processes
results = []
for problem_size in problem_sizes:
results.extend(
_RunSingleIteration(master_vm, problem_size, num_processes,
num_iterations, metadata))
return results
def _CreateMetadataDict(benchmark_spec, model, batch_size, num_gpus):
"""Create metadata dict to be used in run results.
Args:
benchmark_spec: benchmark spec
model: model which was run
batch_size: batch sized used
num_gpus: number of GPUs used
Returns:
metadata dict
"""
vm = benchmark_spec.vms[0]
metadata = dict()
if benchmark_spec.device == GPU:
metadata.update(cuda_toolkit_8.GetMetadata(vm))
metadata['num_gpus'] = num_gpus
metadata['model'] = model
metadata['batch_size'] = batch_size
metadata['forward_only'] = benchmark_spec.forward_only
metadata['data_name'] = benchmark_spec.data_name
metadata['variable_update'] = benchmark_spec.variable_update
metadata['local_parameter_device'] = benchmark_spec.local_parameter_device
metadata['device'] = benchmark_spec.device
metadata['data_format'] = benchmark_spec.data_format
metadata['use_nccl'] = benchmark_spec.use_nccl
metadata['distortions'] = benchmark_spec.distortions
metadata['benchmarks_commit_hash'] = benchmark_spec.benchmarks_commit_hash
metadata['tensorflow_version'] = benchmark_spec.tensorflow_version
return metadata
def Run(benchmark_spec):
"""Sets the GPU clock speed and runs the CUDA PCIe benchmark.
Args:
benchmark_spec: The benchmark specification. Contains all data that is
required to run the benchmark.
Returns:
A list of sample.Sample objects.
"""
vm = benchmark_spec.vms[0]
# Note: The clock speed is set in this function rather than Prepare()
# so that the user can perform multiple runs with a specified
# clock speed without having to re-prepare the VM.
cuda_toolkit_8.SetAndConfirmGpuClocks(vm)
num_iterations = FLAGS.gpu_pcie_bandwidth_iterations
raw_results = []
metadata = {}
metadata.update(cuda_toolkit_8.GetMetadata(vm))
metadata['num_iterations'] = num_iterations
metadata['num_gpus'] = cuda_toolkit_8.QueryNumberOfGpus(vm)
run_command = ('%s/extras/demo_suite/bandwidthTest --device=all' %
cuda_toolkit_8.CUDA_TOOLKIT_INSTALL_DIR)
for i in range(num_iterations):
stdout, _ = vm.RemoteCommand(run_command, should_log=True)
raw_results.append(_ParseOutputFromSingleIteration(stdout))
if 'device_info' not in metadata:
metadata['device_info'] = _ParseDeviceInfo(stdout)
return _CalculateMetricsOverAllIterations(raw_results, metadata)
def _CreateMetadataDict(benchmark_spec):
"""Create metadata dict to be used in run results.
Args:
benchmark_spec: benchmark spec
Returns:
metadata dict
"""
vm = benchmark_spec.vms[0]
metadata = dict()
if benchmark_spec.device == GPU:
metadata.update(cuda_toolkit_8.GetMetadata(vm))
metadata['batch_size'] = benchmark_spec.batch_size
metadata['num_epochs'] = benchmark_spec.num_epochs
metadata['device'] = benchmark_spec.device
metadata['num_layers'] = benchmark_spec.num_layers
metadata['model'] = benchmark_spec.model
metadata['mxnet_version'] = benchmark_spec.mxnet_version
metadata['commit'] = mxnet_cnn.GetCommit(vm)
return metadata
def _CreateMetadataDict(benchmark_spec):
"""Create metadata dict to be used in run results.
Args:
benchmark_spec: benchmark spec
Returns:
metadata dict
"""
vm = benchmark_spec.vms[0]
metadata = dict()
metadata.update(cuda_toolkit_8.GetMetadata(vm))
metadata['num_nodes'] = len(benchmark_spec.vms)
metadata['cpus_per_rank'] = int(benchmark_spec.cpus_per_rank)
metadata['total_gpus'] = int(benchmark_spec.total_gpus)
metadata['benchmark_version'] = BENCHMARK_VERSION
metadata['runtime'] = int(benchmark_spec.hpcg_runtime)
metadata['problem_size'] = '%s,%s,%s' % (
benchmark_spec.hpcg_problem_size[0],
benchmark_spec.hpcg_problem_size[1],
benchmark_spec.hpcg_problem_size[2])
return metadata
