def setup_server(max_number_of_workers=1,
conda_environment='propertyestimator',
worker_memory=5 * unit.gigabyte,
port=8000):
"""Sets up an estimation server which will take advantage of
an LSF based HPC cluster with access to nVidia GPUs.
Parameters
----------
max_number_of_workers: int
The maximum number of workers to adaptively insert into
the queuing system.
conda_environment: str
The name of the conda environment in which the propertyestimator
package is installed.
worker_memory: Quantity
The maximum amount of memory to request per worker.
port: int
The port that the server should listen for estimation requests on.
"""
working_directory = 'working_directory'
storage_directory = 'storage_directory'
# Remove any existing data.
if os.path.isdir(working_directory):
shutil.rmtree(working_directory)
# Request workers with access to a single CPU and CUDA based GPU.
queue_resources = QueueWorkerResources(
number_of_threads=1,
number_of_gpus=1,
preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA,
per_thread_memory_limit=worker_memory,
wallclock_time_limit="05:59")
# Set up extra commands so that each worker has the correct environment
# set up.
worker_script_commands = [
# Load in the correct conda environment.
f'conda activate {conda_environment}',
# Load in CUDA
f'module load cuda/9.2'
]
calculation_backend = DaskLSFBackend(
minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='gpuqueue',
setup_script_commands=worker_script_commands,
adaptive_interval='1000ms')
# Set up a backend to cache simulation data in.
storage_backend = LocalFileStorage(storage_directory)
# Spin up the server object.
PropertyEstimatorServer(calculation_backend=calculation_backend,
storage_backend=storage_backend,
port=port,
working_directory=working_directory)
def test_dask_job_script_creation():
"""Test creating and starting a new dask LSF backend."""
cpu_backend = DaskLSFBackend()
cpu_backend.start()
assert cpu_backend.job_script() is not None
cpu_backend.stop()
def test_lsf_wrapped_function():
available_resources = QueueWorkerResources()
protocols_to_import = [
protocol_class.__module__ + '.' + protocol_class.__qualname__
for protocol_class in available_protocols.values()
]
per_worker_logging = True
gpu_assignments = None
expected_output = 12345
result = DaskLSFBackend._wrapped_function(
dummy_function,
expected_output,
available_resources=available_resources,
available_protocols=protocols_to_import,
per_worker_logging=per_worker_logging,
gpu_assignments=gpu_assignments)
assert expected_output == result
def test_dask_lsf_creation():
"""Test creating and starting a new dask LSF backend."""
cpu_backend = DaskLSFBackend()
cpu_backend.start()
cpu_backend.stop()
gpu_resources = QueueWorkerResources(
1, 1, preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA)
gpu_commands = [
'module load cuda/9.2',
]
gpu_backend = DaskLSFBackend(resources_per_worker=gpu_resources,
queue_name='gpuqueue',
setup_script_commands=gpu_commands)
gpu_backend.start()
gpu_backend.stop()
def setup_server(backend_type=BackendType.LocalCPU,
max_number_of_workers=1,
conda_environment='propertyestimator',
worker_memory=8 * unit.gigabyte,
port=8000):
working_directory = 'working_directory'
storage_directory = 'storage_directory'
# Remove any existing data.
if os.path.isdir(working_directory):
shutil.rmtree(working_directory)
calculation_backend = None
if backend_type == BackendType.LocalCPU:
calculation_backend = DaskLocalCluster(number_of_workers=1)
elif backend_type == BackendType.LocalGPU:
calculation_backend = DaskLocalCluster(
number_of_workers=1,
resources_per_worker=ComputeResources(
1, 1, ComputeResources.GPUToolkit.CUDA))
elif backend_type == BackendType.GPU:
queue_resources = QueueWorkerResources(
number_of_threads=1,
number_of_gpus=1,
preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA,
per_thread_memory_limit=worker_memory,
wallclock_time_limit="05:59")
worker_script_commands = [
f'export OE_LICENSE="/home/boothros/oe_license.txt"',
f'. /home/boothros/miniconda3/etc/profile.d/conda.sh',
f'conda activate {conda_environment}', f'module load cuda/9.2'
]
extra_script_options = ['-m "ls-gpu lt-gpu"']
calculation_backend = DaskLSFBackend(
minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='gpuqueue',
setup_script_commands=worker_script_commands,
extra_script_options=extra_script_options,
adaptive_interval='1000ms')
elif backend_type == BackendType.CPU:
queue_resources = QueueWorkerResources(
number_of_threads=1,
per_thread_memory_limit=worker_memory,
wallclock_time_limit="01:30")
worker_script_commands = [
f'export OE_LICENSE="/home/boothros/oe_license.txt"',
f'. /home/boothros/miniconda3/etc/profile.d/conda.sh',
f'conda activate {conda_environment}',
]
calculation_backend = DaskLSFBackend(
minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='cpuqueue',
setup_script_commands=worker_script_commands,
adaptive_interval='1000ms')
storage_backend = LocalFileStorage(storage_directory)
server.PropertyEstimatorServer(calculation_backend=calculation_backend,
storage_backend=storage_backend,
port=port,
working_directory=working_directory)
def main():
setup_timestamp_logging()
# Load in the force field
force_field_path = 'smirnoff99Frosst-1.1.0.offxml'
force_field_source = SmirnoffForceFieldSource.from_path(force_field_path)
# Create a data set containing three solvation free energies.
data_set = _create_data_set()
# Set up the compute backend which will run the calculations.
working_directory = 'working_directory'
storage_directory = 'storage_directory'
queue_resources = QueueWorkerResources(
number_of_threads=1,
number_of_gpus=1,
preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA,
per_thread_memory_limit=5 * unit.gigabyte,
wallclock_time_limit="05:59")
worker_script_commands = [
'conda activate propertyestimator', 'module load cuda/10.1'
]
calculation_backend = DaskLSFBackend(
minimum_number_of_workers=1,
maximum_number_of_workers=3,
resources_per_worker=queue_resources,
queue_name='gpuqueue',
setup_script_commands=worker_script_commands,
adaptive_interval='1000ms',
adaptive_class=CustomAdaptive)
# Set up a backend to cache simulation data in.
storage_backend = LocalFileStorage(storage_directory)
# Spin up the server object.
PropertyEstimatorServer(calculation_backend=calculation_backend,
storage_backend=storage_backend,
port=8005,
working_directory=working_directory)
# Request the estimates.
property_estimator = client.PropertyEstimatorClient(
client.ConnectionOptions(server_port=8005))
options = PropertyEstimatorOptions()
options.allowed_calculation_layers = ['SimulationLayer']
workflow_options = WorkflowOptions(
WorkflowOptions.ConvergenceMode.NoChecks)
options.workflow_options = {
'SolvationFreeEnergy': {
'SimulationLayer': workflow_options
}
}
options.workflow_schemas = {
'SolvationFreeEnergy': {
'SimulationLayer': _get_fixed_lambda_schema(workflow_options)
}
}
request = property_estimator.request_estimate(
property_set=data_set,
force_field_source=force_field_source,
options=options)
# Wait for the results.
results = request.results(True, 5)
# Save the result to file.
with open('solvation_free_energy_simulation.json', 'wb') as file:
json_results = json.dumps(results,
sort_keys=True,
indent=2,
separators=(',', ': '),
cls=TypedJSONEncoder)
file.write(json_results.encode('utf-8'))
def setup_server(backend_type=BackendType.LocalCPU, max_number_of_workers=1):
"""Creates a new estimation server object.
Parameters
----------
backend_type: BackendType
The type of backend to use.
max_number_of_workers: int
The maximum number of compute workers to spin up.
Returns
-------
PropertyEstimatorServer
The server object.
"""
# Set the name of the directory in which all temporary files
# will be generated.
working_directory = 'working_directory'
# Remove any existing data.
if path.isdir(working_directory):
shutil.rmtree(working_directory)
# Set the name of the directory in which all cached simulation data
# will be stored..
storage_directory = 'storage_directory'
# Set up the backend which will perform any calculations.
calculation_backend = None
if backend_type == BackendType.LocalCPU:
# A backend which will run all calculations on the local machines CPUs.
calculation_backend = DaskLocalCluster(number_of_workers=max_number_of_workers)
if backend_type == BackendType.LocalGPU:
# A backend which will run all calculations on the local machines GPUs.
compute_resources = ComputeResources(number_of_threads=1, number_of_gpus=1)
calculation_backend = DaskLocalCluster(number_of_workers=max_number_of_workers,
resources_per_worker=compute_resources)
elif backend_type == BackendType.LilacGPU:
# A backend which will run all calculations on the MSKCC `lilac` cluster, taking
# advantage of the available GPUs.
queue_resources = QueueWorkerResources(number_of_threads=1,
number_of_gpus=1,
preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA,
per_thread_memory_limit=5 * unit.gigabyte,
wallclock_time_limit="05:59")
extra_script_options = [
'-m "ls-gpu lt-gpu"'
]
worker_script_commands = [
'export OE_LICENSE="/home/boothros/oe_license.txt"',
'. /home/boothros/miniconda3/etc/profile.d/conda.sh',
'conda activate forcebalance',
'module load cuda/9.2'
]
calculation_backend = DaskLSFBackend(minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='gpuqueue',
setup_script_commands=worker_script_commands,
extra_script_options=extra_script_options,
adaptive_interval='1000ms')
elif backend_type == BackendType.LilacCPU:
# A backend which will run all calculations on the MSKCC `lilac` cluster using onlu
# CPUs.
queue_resources = QueueWorkerResources(number_of_threads=1,
per_thread_memory_limit=5 * unit.gigabyte,
wallclock_time_limit="01:30")
worker_script_commands = [
'export OE_LICENSE="/home/boothros/oe_license.txt"',
'. /home/boothros/miniconda3/etc/profile.d/conda.sh',
'conda activate forcebalance',
]
calculation_backend = DaskLSFBackend(minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='cpuqueue',
setup_script_commands=worker_script_commands,
adaptive_interval='1000ms')
# Set up the storage backend.
storage_backend = LocalFileStorage(storage_directory)
# Set up the server itself.
server = PropertyEstimatorServer(calculation_backend=calculation_backend,
storage_backend=storage_backend,
working_directory=working_directory)
return server
def setup_server(backend_type=BackendType.LocalCPU,
max_number_of_workers=1,
conda_environment='propertyestimator',
worker_memory=4 * unit.gigabyte,
port=8000,
cuda_version='10.1'):
"""A convenience function to sets up an estimation server which will can advantage
of different compute backends.
Parameters
----------
backend_type: BackendType
The type of compute backend to use.
max_number_of_workers: int
The maximum number of workers to adaptively insert into
the queuing system.
conda_environment: str
The name of the conda environment in which the propertyestimator
package is installed.
worker_memory: Quantity
The maximum amount of memory to request per worker.
port: int
The port that the server should listen for estimation requests on.
cuda_version: str
The version of CUDA to use if running on a backend which supports
GPUs.
"""
working_directory = 'working_directory'
storage_directory = 'storage_directory'
# Remove any existing data.
if os.path.isdir(working_directory):
shutil.rmtree(working_directory)
calculation_backend = None
if backend_type == BackendType.LocalCPU:
calculation_backend = DaskLocalCluster(
number_of_workers=max_number_of_workers)
elif backend_type == BackendType.LocalGPU:
calculation_backend = DaskLocalCluster(
number_of_workers=max_number_of_workers,
resources_per_worker=ComputeResources(
1, 1, ComputeResources.GPUToolkit.CUDA))
elif backend_type == BackendType.GPU:
queue_resources = QueueWorkerResources(
number_of_threads=1,
number_of_gpus=1,
preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA,
per_thread_memory_limit=worker_memory,
wallclock_time_limit="05:59")
worker_script_commands = [
f'conda activate {conda_environment}',
f'module load cuda/{cuda_version}'
]
calculation_backend = DaskLSFBackend(
minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='gpuqueue',
setup_script_commands=worker_script_commands,
adaptive_interval='1000ms')
elif backend_type == BackendType.CPU:
queue_resources = QueueWorkerResources(
number_of_threads=1,
per_thread_memory_limit=worker_memory,
wallclock_time_limit="01:30")
worker_script_commands = [f'conda activate {conda_environment}']
calculation_backend = DaskLSFBackend(
minimum_number_of_workers=1,
maximum_number_of_workers=max_number_of_workers,
resources_per_worker=queue_resources,
queue_name='cpuqueue',
setup_script_commands=worker_script_commands,
adaptive_interval='1000ms')
# Set up a backend to cache simulation data in.
storage_backend = LocalFileStorage(storage_directory)
# Spin up the server object.
server.PropertyEstimatorServer(calculation_backend=calculation_backend,
storage_backend=storage_backend,
port=port,
working_directory=working_directory)