Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[doc] Update Dask deployment options (#9008)

Browse Source
This commit is contained in:
Sarah Charlotte Johnson 2023-04-06 10:09:15 -07:00 committed by GitHub
parent 1cf4d93246
commit ebd64f6e22
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 5 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
doc/tutorials/dask.rst
View File

@ -66,7 +66,7 @@ on a dask cluster:
Here we first create a cluster in single-node mode with Here we first create a cluster in single-node mode with
:py:class:`distributed.LocalCluster`, then connect a :py:class:`distributed.Client` to :py:class:`distributed.LocalCluster`, then connect a :py:class:`distributed.Client` to
this cluster, setting up an environment for later computation. Notice that the cluster this cluster, setting up an environment for later computation. Notice that the cluster
construction is guared by ``__name__ == "__main__"``, which is necessary otherwise there construction is guarded by ``__name__ == "__main__"``, which is necessary otherwise there
might be obscure errors. might be obscure errors.
We then create a :py:class:`xgboost.dask.DaskDMatrix` object and pass it to We then create a :py:class:`xgboost.dask.DaskDMatrix` object and pass it to
@ -226,13 +226,9 @@ collection.
Working with other clusters Working with other clusters
*************************** ***************************
``LocalCluster`` is mostly used for testing. In real world applications some other Using Dask's ``LocalCluster`` is convenient for getting started quickly on a single-machine. Once you're ready to scale your work, though, there are a number of ways to deploy Dask on a distributed cluster. You can use `Dask-CUDA <https://docs.rapids.ai/api/dask-cuda/stable/quickstart.html>`_, for example, for GPUs and you can use Dask Cloud Provider to `deploy Dask clusters in the cloud <https://docs.dask.org/en/stable/deploying.html#cloud>`_. See the `Dask documentation for a more comprehensive list <https://docs.dask.org/en/stable/deploying.html#distributed-computing>`_.
clusters might be preferred. Examples are like ``LocalCUDACluster`` for single node
multi-GPU instance, manually launched cluster by using command line utilities like In the example below, a ``KubeCluster`` is used for `deploying Dask on Kubernetes <https://docs.dask.org/en/stable/deploying-kubernetes.html>`_:
``dask-worker`` from ``distributed`` for not yet automated environments. Some special
clusters like ``KubeCluster`` from ``dask-kubernetes`` package are also possible. The
dask API in xgboost is orthogonal to the cluster type and can be used with any of them. A
typical testing workflow with ``KubeCluster`` looks like this:
.. code-block:: python .. code-block:: python
@ -272,8 +268,7 @@ typical testing workflow with ``KubeCluster`` looks like this:
# main function will connect to that cluster and start training xgboost model. # main function will connect to that cluster and start training xgboost model.
main() main()
Different cluster classes might have subtle differences like network configuration, or
However, these clusters might have their subtle differences like network configuration, or
specific cluster implementation might contains bugs that we are not aware of. Open an specific cluster implementation might contains bugs that we are not aware of. Open an
issue if such case is found and there's no documentation on how to resolve it in that issue if such case is found and there's no documentation on how to resolve it in that
cluster implementation. cluster implementation.