DaskLocalCluster backend wraps around the dask LocalCluster class to distribute tasks on a single machine:
worker_resources = ComputeResources( number_of_threads=1, number_of_gpus=1, preferred_gpu_toolkit=ComputeResources.GPUToolkit.CUDA, ) with DaskLocalCluster(number_of_workers=1, resources_per_worker=worker_resources) as local_backend: local_backend.submit_task(logging.info, "Hello World") ...
Its main purpose is for use when debugging calculations locally, or when running calculations on machines with large numbers of CPUs or GPUs.
DaskSLURMBackend backends wrap around the dask LSFCluster, PBSCluster, and SLURMCluster classes
respectively, and both inherit the
BaseDaskJobQueueBackend class which implements the core of their
functionality. They predominantly run in an adaptive mode, whereby the backend will automatically scale up or down
the number of workers based on the current number of tasks that the backend is trying to execute.
These backends integrate with the queueing systems which most HPC cluster use to manage task execution. They work by submitting jobs into the queueing system which themselves spawn dask workers, which in turn then execute tasks on the available compute nodes:
# Create the object which describes the compute resources each worker should request from # the queueing system. worker_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", ) # Create the backend object. setup_script_commands = [ f"conda activate evaluator", f"module load cuda/10.1", ] 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=setup_script_commands, ) # Perform some tasks. with calculation_backend: calculation_backend.submit_task(logging.info, "Hello World") ...
setup_script_commands argument takes a list of commands which should be run by the queue job submission
script before spawning the actual worker. This enables setting up custom environments, and setting any required
To ensure optimal behaviour we recommend changing / uncommenting the following settings in the dask distributed
configuration file (this can be found at
distributed: worker: daemon: False comm: timeouts: connect: 10s tcp: 30s deploy: lost-worker-timeout: 15s
See the dask documentation for more information about changing
The calculation backends alos allows the user to specify the GPU platform and precision level. Users can specify
OpenCL as the preferred_gpu_toolkit using the
GPUToolkit enum class. The
default precision level is set to
mixed but can be overridden by specifying preferred_gpu_precision using the
GPUPrecision enum class:
worker_resources = ComputeResources( number_of_threads=1, number_of_gpus=1, preferred_gpu_toolkit=ComputeResources.GPUToolkit.OpenCL, preferred_gpu_precision=ComputeResources.GPUPrecision.mixed, )
GPUToolkit.auto, the framework will determine the fastest available platform based on the precision level:
worker_resources = ComputeResources( number_of_threads=1, number_of_gpus=1, preferred_gpu_toolkit=ComputeResources.GPUToolkit.auto, preferred_gpu_precision=ComputeResources.GPUPrecision.mixed, )