Distributed Learning Guide

This guide describes distributed learning in LightGBM. Distributed learning allows the use of multiple machines to produce a single model.

Follow the Quick Start to know how to use LightGBM first.

How Distributed LightGBM Works

This section describes how distributed learning in LightGBM works. To learn how to do this in various programming languages and frameworks, please see Integrations.

Choose Appropriate Parallel Algorithm

LightGBM provides 3 distributed learning algorithms now.

Parallel Algorithm

How to Use

Data parallel

tree_learner=data

Feature parallel

tree_learner=feature

Voting parallel

tree_learner=voting

These algorithms are suited for different scenarios, which is listed in the following table:

#data is small

#data is large

#feature is small

Feature Parallel

Data Parallel

#feature is large

Feature Parallel

Voting Parallel

More details about these parallel algorithms can be found in optimization in distributed learning.

Integrations

This section describes how to run distributed LightGBM training in various programming languages and frameworks. To learn how distributed learning in LightGBM works generally, please see How Distributed LightGBM Works.

Apache Spark

Apache Spark users can use SynapseML for machine learning workflows with LightGBM. This project is not maintained by LightGBM’s maintainers.

See this SynapseML example and the SynapseML documentation for additional information on using LightGBM on Spark.

Note

SynapseML is not maintained by LightGBM’s maintainers. Bug reports or feature requests should be directed to https://github.com/microsoft/SynapseML/issues.

Dask

New in version 3.2.0.

LightGBM’s Python package supports distributed learning via Dask. This integration is maintained by LightGBM’s maintainers.

Warning

Dask integration is only tested on Linux.

Dask Examples

For sample code using lightgbm.dask, see these Dask examples.

Training with Dask

This section contains detailed information on performing LightGBM distributed training using Dask.

Configuring the Dask Cluster

Allocating Threads

When setting up a Dask cluster for training, give each Dask worker process at least two threads. If you do not do this, training might be substantially slower because communication work and training work will block each other.

If you do not have other significant processes competing with Dask for resources, just accept the default nthreads from your chosen dask.distributed cluster.

from distributed import Client, LocalCluster

cluster = LocalCluster(n_workers=3)
client = Client(cluster)

Managing Memory

Use the Dask diagnostic dashboard or your preferred monitoring tool to monitor Dask workers’ memory consumption during training. As described in the Dask worker documentation, Dask workers will automatically start spilling data to disk if memory consumption gets too high. This can substantially slow down computations, since disk I/O is usually much slower than reading the same data from memory.

At 60% of memory load, [Dask will] spill least recently used data to disk

To reduce the risk of hitting memory limits, consider restarting each worker process before running any data loading or training code.

client.restart()
Setting Up Training Data

The estimators in lightgbm.dask expect that matrix-like or array-like data are provided in Dask DataFrame, Dask Array, or (in some cases) Dask Series format. See the Dask DataFrame documentation and the Dask Array documentation for more information on how to create such data structures.

On the left, rectangles showing a 5 by 5 grid for a local dataset. On the right, two circles representing Dask workers, one with a 3 by 5 grid and one with a 2 by 5 grid.

While setting up for training, lightgbm will concatenate all of the partitions on a worker into a single dataset. Distributed training then proceeds with one LightGBM worker process per Dask worker.

A section labeled "before" showing two grids and a section labeled "after" showing a single grid that looks like the two from "before" stacked one on top of the other.

When setting up data partitioning for LightGBM training with Dask, try to follow these suggestions:

  • ensure that each worker in the cluster has some of the training data

  • try to give each worker roughly the same amount of data, especially if your dataset is small

  • if you plan to train multiple models (for example, to tune hyperparameters) on the same data, use client.persist() before training to materialize the data one time

Using a Specific Dask Client

In most situations, you should not need to tell lightgbm.dask to use a specific Dask client. By default, the client returned by distributed.default_client() will be used.

However, you might want to explicitly control the Dask client used by LightGBM if you have multiple active clients in the same session. This is useful in more complex workflows like running multiple training jobs on different Dask clusters.

LightGBM’s Dask estimators support setting an attribute client to control the client that is used.

import lightgbm as lgb
from distributed import Client, LocalCluster

cluster = LocalCluster()
client = Client(cluster)

# option 1: keyword argument in constructor
dask_model = lgb.DaskLGBMClassifier(client=client)

# option 2: set_params() after construction
dask_model = lgb.DaskLGBMClassifier()
dask_model.set_params(client=client)
Using Specific Ports

At the beginning of training, lightgbm.dask sets up a LightGBM network where each Dask worker runs one long-running task that acts as a LightGBM worker. During training, LightGBM workers communicate with each other over TCP sockets. By default, random open ports are used when creating these sockets.

If the communication between Dask workers in the cluster used for training is restricted by firewall rules, you must tell LightGBM exactly what ports to use.

Option 1: provide a specific list of addresses and ports

LightGBM supports a parameter machines, a comma-delimited string where each entry refers to one worker (host name or IP) and a port that that worker will accept connections on. If you provide this parameter to the estimators in lightgbm.dask, LightGBM will not search randomly for ports.

For example, consider the case where you are running one Dask worker process on each of the following IP addresses:

10.0.1.0
10.0.2.0
10.0.3.0

You could edit your firewall rules to allow traffic on one additional port on each of these hosts, then provide machines directly.

import lightgbm as lgb

machines = "10.0.1.0:12401,10.0.2.0:12402,10.0.3.0:15000"
dask_model = lgb.DaskLGBMRegressor(machines=machines)

If you are running multiple Dask worker processes on physical host in the cluster, be sure that there are multiple entries for that IP address, with different ports. For example, if you were running a cluster with nprocs=2 (2 Dask worker processes per machine), you might open two additional ports on each of these hosts, then provide machines as follows.

import lightgbm as lgb

machines = ",".join([
  "10.0.1.0:16000",
  "10.0.1.0:16001",
  "10.0.2.0:16000",
  "10.0.2.0:16001",
])
dask_model = lgb.DaskLGBMRegressor(machines=machines)

Warning

Providing machines gives you complete control over the networking details of training, but it also makes the training process fragile. Training will fail if you use machines and any of the following are true:

  • any of the ports mentioned in machines are not open when training begins

  • some partitions of the training data are held by machines that that are not present in machines

  • some machines mentioned in machines do not hold any of the training data

Option 2: specify one port to use on every worker

If you are only running one Dask worker process on each host, and if you can reliably identify a port that is open on every host, using machines is unnecessarily complicated. If local_listen_port is given and machines is not, LightGBM will not search for ports randomly, but it will limit the list of addresses in the LightGBM network to those Dask workers that have a piece of the training data.

For example, consider the case where you are running one Dask worker process on each of the following IP addresses:

10.0.1.0
10.0.2.0
10.0.3.0

You could edit your firewall rules to allow communication between any of the workers over one port, then provide that port via parameter local_listen_port.

import lightgbm as lgb

dask_model = lgb.DaskLGBMRegressor(local_listen_port=12400)

Warning

Providing local_listen_port is slightly less fragile than machines because LightGBM will automatically figure out which workers have pieces of the training data. However, using this method, training can fail if any of the following are true:

  • the port local_listen_port is not open on any of the worker hosts

  • any machine has multiple Dask worker processes running on it

Prediction with Dask

The estimators from lightgbm.dask can be used to create predictions based on data stored in Dask collections. In that interface, .predict() expects a Dask Array or Dask DataFrame, and returns a Dask Array of predictions.

See the Dask prediction example for some sample code that shows how to perform Dask-based prediction.

For model evaluation, consider using the metrics functions from dask-ml. Those functions are intended to provide the same API as equivalent functions in sklearn.metrics, but they use distributed computation powered by Dask to compute metrics without all of the input data ever needing to be on a single machine.

Saving Dask Models

After training with Dask, you have several options for saving a fitted model.

Option 1: pickle the Dask estimator

LightGBM’s Dask estimators can be pickled directly with cloudpickle, joblib, or pickle.

import dask.array as da
import pickle
import lightgbm as lgb
from distributed import Client, LocalCluster

cluster = LocalCluster(n_workers=2)
client = Client(cluster)

X = da.random.random((1000, 10), (500, 10))
y = da.random.random((1000,), (500,))

dask_model = lgb.DaskLGBMRegressor()
dask_model.fit(X, y)

with open("dask-model.pkl", "wb") as f:
    pickle.dump(dask_model, f)

A model saved this way can then later be loaded with whichever serialization library you used to save it.

import pickle
with open("dask-model.pkl", "rb") as f:
    dask_model = pickle.load(f)

Note

If you explicitly set a Dask client (see Using a Specific Dask Client), it will not be saved when pickling the estimator. When loading a Dask estimator from disk, if you need to use a specific client you can add it after loading with dask_model.set_params(client=client).

Option 2: pickle the sklearn estimator

The estimators available from lightgbm.dask can be converted to an instance of the equivalent class from lightgbm.sklearn. Choosing this option allows you to use Dask for training but avoid depending on any Dask libraries at scoring time.

import dask.array as da
import joblib
import lightgbm as lgb
from distributed import Client, LocalCluster

cluster = LocalCluster(n_workers=2)
client = Client(cluster)

X = da.random.random((1000, 10), (500, 10))
y = da.random.random((1000,), (500,))

dask_model = lgb.DaskLGBMRegressor()
dask_model.fit(X, y)

# convert to sklearn equivalent
sklearn_model = dask_model.to_local()

print(type(sklearn_model))
#> lightgbm.sklearn.LGBMRegressor

joblib.dump(sklearn_model, "sklearn-model.joblib")

A model saved this way can then later be loaded with whichever serialization library you used to save it.

import joblib

sklearn_model = joblib.load("sklearn-model.joblib")

Option 3: save the LightGBM Booster

The lowest-level model object in LightGBM is the lightgbm.Booster. After training, you can extract a Booster from the Dask estimator.

import dask.array as da
import lightgbm as lgb
from distributed import Client, LocalCluster

cluster = LocalCluster(n_workers=2)
client = Client(cluster)

X = da.random.random((1000, 10), (500, 10))
y = da.random.random((1000,), (500,))

dask_model = lgb.DaskLGBMRegressor()
dask_model.fit(X, y)

# get underlying Booster object
bst = dask_model.booster_

From the point forward, you can use any of the following methods to save the Booster:

  • serialize with cloudpickle, joblib, or pickle

  • bst.dump_model(): dump the model to a dictionary which could be written out as JSON

  • bst.model_to_string(): dump the model to a string in memory

  • bst.save_model(): write the output of bst.model_to_string() to a text file

Kubeflow

Kubeflow Fairing supports LightGBM distributed training. These examples show how to get started with LightGBM and Kubeflow Fairing in a hybrid cloud environment.

Kubeflow users can also use the Kubeflow XGBoost Operator for machine learning workflows with LightGBM. You can see this example for more details.

Kubeflow integrations for LightGBM are not maintained by LightGBM’s maintainers.

Note

The Kubeflow integrations for LightGBM are not maintained by LightGBM’s maintainers. Bug reports or feature requests should be directed to https://github.com/kubeflow/fairing/issues or https://github.com/kubeflow/xgboost-operator/issues.

LightGBM CLI

Preparation

By default, distributed learning with LightGBM uses socket-based communication.

If you need to build distributed version with MPI support, please refer to Installation Guide.

Socket Version

It needs to collect IP of all machines that want to run distributed learning in and allocate one TCP port (assume 12345 here) for all machines, and change firewall rules to allow income of this port (12345). Then write these IP and ports in one file (assume mlist.txt), like following:

machine1_ip 12345
machine2_ip 12345
MPI Version

It needs to collect IP (or hostname) of all machines that want to run distributed learning in. Then write these IP in one file (assume mlist.txt) like following:

machine1_ip
machine2_ip

Note: For Windows users, need to start “smpd” to start MPI service. More details can be found here.

Run Distributed Learning

Socket Version
  1. Edit following parameters in config file:

    tree_learner=your_parallel_algorithm, edit your_parallel_algorithm (e.g. feature/data) here.

    num_machines=your_num_machines, edit your_num_machines (e.g. 4) here.

    machine_list_file=mlist.txt, mlist.txt is created in Preparation section.

    local_listen_port=12345, 12345 is allocated in Preparation section.

  2. Copy data file, executable file, config file and mlist.txt to all machines.

  3. Run following command on all machines, you need to change your_config_file to real config file.

    For Windows: lightgbm.exe config=your_config_file

    For Linux: ./lightgbm config=your_config_file

MPI Version
  1. Edit following parameters in config file:

    tree_learner=your_parallel_algorithm, edit your_parallel_algorithm (e.g. feature/data) here.

    num_machines=your_num_machines, edit your_num_machines (e.g. 4) here.

  2. Copy data file, executable file, config file and mlist.txt to all machines.

    Note: MPI needs to be run in the same path on all machines.

  3. Run following command on one machine (not need to run on all machines), need to change your_config_file to real config file.

    For Windows:

    mpiexec.exe /machinefile mlist.txt lightgbm.exe config=your_config_file
    

    For Linux:

    mpiexec --machinefile mlist.txt ./lightgbm config=your_config_file
    

Example

Ray

Ray is a Python-based framework for distributed computing. The lightgbm_ray project, maintained within the official Ray GitHub organization, can be used to perform distributed LightGBM training using ray.

See the lightgbm_ray documentation for usage examples.

Note

lightgbm_ray is not maintained by LightGBM’s maintainers. Bug reports or feature requests should be directed to https://github.com/ray-project/lightgbm_ray/issues.

Mars

Mars is a tensor-based framework for large-scale data computation. LightGBM integration, maintained within the Mars GitHub repository, can be used to perform distributed LightGBM training using pymars.

See the mars documentation for usage examples.

Note

Mars is not maintained by LightGBM’s maintainers. Bug reports or feature requests should be directed to https://github.com/mars-project/mars/issues.