Add support for configuring Dask distributed (#2049)

Co-authored-by: Valeriu Predoi <valeriu.predoi@gmail.com>
Co-authored-by: Rémi Kazeroni <remi.kazeroni@dlr.de>

Author: 3 people

doc/conf.py

@@ -423,6 +423,8 @@
     (f'https://docs.esmvaltool.org/projects/ESMValCore/en/{rtd_version}/',
      None),
     'esmvaltool': (f'https://docs.esmvaltool.org/en/{rtd_version}/', None),
+    'dask': ('https://docs.dask.org/en/stable/', None),
+    'distributed': ('https://distributed.dask.org/en/stable/', None),
     'iris': ('https://scitools-iris.readthedocs.io/en/latest/', None),
     'iris-esmf-regrid': ('https://iris-esmf-regrid.readthedocs.io/en/latest',
                          None),

doc/quickstart/configure.rst

@@ -199,6 +199,161 @@ the user.
    debugging, etc. You can even provide any config user value as a run flag
    ``--argument_name argument_value``
 
+.. _config-dask:
+
+Dask distributed configuration
+==============================
+
+The :ref:`preprocessor functions <preprocessor_functions>` and many of the
+:ref:`Python diagnostics in ESMValTool <esmvaltool:recipes>` make use of the
+:ref:`Iris <iris:iris_docs>` library to work with the data.
+In Iris, data can be either :ref:`real or lazy <iris:real_and_lazy_data>`.
+Lazy data is represented by `dask arrays <https://docs.dask.org/en/stable/array.html>`_.
+Dask arrays consist of many small
+`numpy arrays <https://numpy.org/doc/stable/user/absolute_beginners.html#what-is-an-array>`_
+(called chunks) and if possible, computations are run on those small arrays in
+parallel.
+In order to figure out what needs to be computed when, Dask makes use of a
+'`scheduler <https://docs.dask.org/en/stable/scheduling.html>`_'.
+The default scheduler in Dask is rather basic, so it can only run on a single
+computer and it may not always find the optimal task scheduling solution,
+resulting in excessive memory use when using e.g. the
+:func:`esmvalcore.preprocessor.multi_model_statistics` preprocessor function.
+Therefore it is recommended that you take a moment to configure the
+`Dask distributed <https://distributed.dask.org>`_ scheduler.
+A Dask scheduler and the 'workers' running the actual computations, are
+collectively called a 'Dask cluster'.
+
+In ESMValCore, the Dask cluster can configured by creating a file called
+``~/.esmvaltool/dask.yml``, where ``~`` is short for your home directory.
+In this file, under the ``client`` keyword, the arguments to
+:obj:`distributed.Client` can be provided.
+Under the ``cluster`` keyword, the type of cluster (e.g.
+:obj:`distributed.LocalCluster`), as well as any arguments required to start
+the cluster can be provided.
+Extensive documentation on setting up Dask Clusters is available
+`here <https://docs.dask.org/en/latest/deploying.html>`__.
+
+.. warning::
+
+  The format of the ``~/.esmvaltool/dask.yml`` configuration file is not yet
+  fixed and may change in the next release of ESMValCore.
+
+.. note::
+
+  If not all preprocessor functions support lazy data, computational
+  performance may be best with the default scheduler.
+  See `issue #674 <https://github.com/ESMValGroup/ESMValCore/issues/674>`_ for
+  progress on making all preprocessor functions lazy.
+
+**Example configurations**
+
+*Personal computer*
+
+Create a Dask distributed cluster on the computer running ESMValCore using
+all available resources:
+
+.. code:: yaml
+
+  cluster:
+    type: distributed.LocalCluster
+
+this should work well for most personal computers.
+
+.. note::
+
+   Note that, if running this configuration on a shared node of an HPC cluster,
+   Dask will try and use as many resources it can find available, and this may
+   lead to overcrowding the node by a single user (you)!
+
+*Shared computer*
+
+Create a Dask distributed cluster on the computer running ESMValCore, with
+2 workers with 4 threads/4 GiB of memory each (8 GiB in total):
+
+.. code:: yaml
+
+  cluster:
+    type: distributed.LocalCluster
+    n_workers: 2
+    threads_per_worker: 4
+    memory_limit: 4 GiB
+
+this should work well for shared computers.
+
+*Computer cluster*
+
+Create a Dask distributed cluster on the
+`Levante <https://docs.dkrz.de/doc/levante/running-jobs/index.html>`_
+supercomputer using the `Dask-Jobqueue <https://jobqueue.dask.org/en/latest/>`_
+package:
+
+.. code:: yaml
+
+  cluster:
+    type: dask_jobqueue.SLURMCluster
+    queue: shared
+    account: bk1088
+    cores: 8
+    memory: 7680MiB
+    processes: 2
+    interface: ib0
+    local_directory: "/scratch/b/b381141/dask-tmp"
+    n_workers: 24
+
+This will start 24 workers with ``cores / processes = 4`` threads each,
+resulting in ``n_workers / processes = 12`` Slurm jobs, where each Slurm job
+will request 8 CPU cores and 7680 MiB of memory and start ``processes = 2``
+workers.
+This example will use the fast infiniband network connection (called ``ib0``
+on Levante) for communication between workers running on different nodes.
+It is
+`important to set the right location for temporary storage <https://docs.dask.org/en/latest/deploying-hpc.html#local-storage>`__,
+in this case the ``/scratch`` space is used.
+It is also possible to use environmental variables to configure the temporary
+storage location, if you cluster provides these.
+
+A configuration like this should work well for larger computations where it is
+advantageous to use multiple nodes in a compute cluster.
+See
+`Deploying Dask Clusters on High Performance Computers <https://docs.dask.org/en/latest/deploying-hpc.html>`_
+for more information.
+
+*Externally managed Dask cluster*
+
+Use an externally managed cluster, e.g. a cluster that you started using the
+`Dask Jupyterlab extension <https://github.com/dask/dask-labextension#dask-jupyterlab-extension>`_:
+
+.. code:: yaml
+
+  client:
+    address: '127.0.0.1:8786'
+
+See `here <https://jobqueue.dask.org/en/latest/interactive.html>`_
+for an example of how to configure this on a remote system.
+
+For debugging purposes, it can be useful to start the cluster outside of
+ESMValCore because then
+`Dask dashboard <https://docs.dask.org/en/stable/dashboard.html>`_ remains
+available after ESMValCore has finished running.
+
+**Advice on choosing performant configurations**
+
+The threads within a single worker can access the same memory locations, so
+they may freely pass around chunks, while communicating a chunk between workers
+is done by copying it, so this is (a bit) slower.
+Therefore it is beneficial for performance to have multiple threads per worker.
+However, due to limitations in the CPython implementation (known as the Global
+Interpreter Lock or GIL), only a single thread in a worker can execute Python
+code (this limitation does not apply to compiled code called by Python code,
+e.g. numpy), therefore the best performing configurations will typically not
+use much more than 10 threads per worker.
+
+Due to limitations of the NetCDF library (it is not thread-safe), only one
+of the threads in a worker can read or write to a NetCDF file at a time.
+Therefore, it may be beneficial to use fewer threads per worker if the
+computation is very simple and the runtime is determined by the
+speed with which the data can be read from and/or written to disk.
 
 .. _config-esgf:
 

environment.yml

@@ -10,6 +10,8 @@ dependencies:
   - cftime
   - compilers
   - dask
+  - dask-jobqueue
+  - distributed
   - esgf-pyclient>=0.3.1
   - esmpy!=8.1.0
   - filelock
@@ -18,7 +20,7 @@ dependencies:
   - geopy
   - humanfriendly
   - importlib_resources
-  - iris>=3.4.0
+  - iris>=3.6.0
   - iris-esmf-regrid >=0.6.0  # to work with latest esmpy
   - isodate
   - jinja2

esmvalcore/_main.py

@@ -74,6 +74,7 @@ def process_recipe(recipe_file: Path, session):
     import shutil
 
     from esmvalcore._recipe.recipe import read_recipe_file
+    from esmvalcore.config._dask import check_distributed_config
     if not recipe_file.is_file():
         import errno
         raise OSError(errno.ENOENT, "Specified recipe file does not exist",
@@ -103,6 +104,8 @@ def process_recipe(recipe_file: Path, session):
     logger.info("If you experience memory problems, try reducing "
                 "'max_parallel_tasks' in your user configuration file.")
 
+    check_distributed_config()
+
     if session['compress_netcdf']:
         logger.warning(
             "You have enabled NetCDF compression. Accessing .nc files can be "

esmvalcore/_task.py

@@ -19,9 +19,11 @@
 
 import psutil
 import yaml
+from distributed import Client
 
 from ._citation import _write_citation_files
 from ._provenance import TrackedFile, get_task_provenance
+from .config._dask import get_distributed_client
 from .config._diagnostics import DIAGNOSTICS, TAGS
 
 
@@ -718,10 +720,22 @@ def run(self, max_parallel_tasks: Optional[int] = None) -> None:
         max_parallel_tasks : int
             Number of processes to run. If `1`, run the tasks sequentially.
         """
-        if max_parallel_tasks == 1:
-            self._run_sequential()
-        else:
-            self._run_parallel(max_parallel_tasks)
+        with get_distributed_client() as client:
+            if client is None:
+                address = None
+            else:
+                address = client.scheduler.address
+                for task in self.flatten():
+                    if (isinstance(task, DiagnosticTask)
+                            and Path(task.script).suffix.lower() == '.py'):
+                        # Only insert the scheduler address if running a
+                        # Python script.
+                        task.settings['scheduler_address'] = address
+
+            if max_parallel_tasks == 1:
+                self._run_sequential()
+            else:
+                self._run_parallel(address, max_parallel_tasks)
 
     def _run_sequential(self) -> None:
         """Run tasks sequentially."""
@@ -732,7 +746,7 @@ def _run_sequential(self) -> None:
         for task in sorted(tasks, key=lambda t: t.priority):
             task.run()
 
-    def _run_parallel(self, max_parallel_tasks=None):
+    def _run_parallel(self, scheduler_address, max_parallel_tasks):
         """Run tasks in parallel."""
         scheduled = self.flatten()
         running = {}
@@ -757,7 +771,8 @@ def done(task):
                     if len(running) >= max_parallel_tasks:
                         break
                     if all(done(t) for t in task.ancestors):
-                        future = pool.apply_async(_run_task, [task])
+                        future = pool.apply_async(_run_task,
+                                                  [task, scheduler_address])
                         running[task] = future
                         scheduled.remove(task)
 
@@ -790,7 +805,14 @@ def _copy_results(task, future):
     task.output_files, task.products = future.get()
 
 
-def _run_task(task):
+def _run_task(task, scheduler_address):
     """Run task and return the result."""
-    output_files = task.run()
+    if scheduler_address is None:
+        client = contextlib.nullcontext()
+    else:
+        client = Client(scheduler_address)
+
+    with client:
+        output_files = task.run()
+
     return output_files, task.products

esmvalcore/config/_dask.py

@@ -0,0 +1,79 @@
+"""Configuration for Dask distributed."""
+import contextlib
+import importlib
+import logging
+from pathlib import Path
+
+import yaml
+from distributed import Client
+
+logger = logging.getLogger(__name__)
+
+CONFIG_FILE = Path.home() / '.esmvaltool' / 'dask.yml'
+
+
+def check_distributed_config():
+    """Check the Dask distributed configuration."""
+    if not CONFIG_FILE.exists():
+        logger.warning(
+            "Using the Dask basic scheduler. This may lead to slow "
+            "computations and out-of-memory errors. "
+            "Note that the basic scheduler may still be the best choice for "
+            "preprocessor functions that are not lazy. "
+            "In that case, you can safely ignore this warning. "
+            "See https://docs.esmvaltool.org/projects/ESMValCore/en/latest/"
+            "quickstart/configure.html#dask-distributed-configuration for "
+            "more information. ")
+
+
+@contextlib.contextmanager
+def get_distributed_client():
+    """Get a Dask distributed client."""
+    dask_args = {}
+    if CONFIG_FILE.exists():
+        config = yaml.safe_load(CONFIG_FILE.read_text(encoding='utf-8'))
+        if config is not None:
+            dask_args = config
+
+    client_args = dask_args.get('client') or {}
+    cluster_args = dask_args.get('cluster') or {}
+
+    # Start a cluster, if requested
+    if 'address' in client_args:
+        # Use an externally managed cluster.
+        cluster = None
+        if cluster_args:
+            logger.warning(
+                "Not using Dask 'cluster' settings from %s because a cluster "
+                "'address' is already provided in 'client'.", CONFIG_FILE)
+    elif cluster_args:
+        # Start cluster.
+        cluster_type = cluster_args.pop(
+            'type',
+            'distributed.LocalCluster',
+        )
+        cluster_module_name, cluster_cls_name = cluster_type.rsplit('.', 1)
+        cluster_module = importlib.import_module(cluster_module_name)
+        cluster_cls = getattr(cluster_module, cluster_cls_name)
+        cluster = cluster_cls(**cluster_args)
+        client_args['address'] = cluster.scheduler_address
+    else:
+        # No cluster configured, use Dask basic scheduler, or a LocalCluster
+        # managed through Client.
+        cluster = None
+
+    # Start a client, if requested
+    if dask_args:
+        client = Client(**client_args)
+        logger.info("Dask dashboard: %s", client.dashboard_link)
+    else:
+        logger.info("Using the Dask basic scheduler.")
+        client = None
+
+    try:
+        yield client
+    finally:
+        if client is not None:
+            client.close()
+        if cluster is not None:
+            cluster.close()

esmvalcore/experimental/recipe.py

@@ -10,7 +10,7 @@
 import yaml
 
 from esmvalcore._recipe.recipe import Recipe as RecipeEngine
-from esmvalcore.config import CFG, Session
+from esmvalcore.config import CFG, Session, _dask
 
 from ._logging import log_to_dir
 from .recipe_info import RecipeInfo
@@ -132,6 +132,7 @@ def run(
             session['diagnostics'] = task
 
         with log_to_dir(session.run_dir):
+            _dask.check_distributed_config()
             self._engine = self._load(session=session)
             self._engine.run()