Distributed XGBoost on Ray
XGBoost-Ray is a distributed backend for XGBoost, built on top of distributed computing framework Ray.
XGBoost-Ray
- enables multi-node and multi-GPU training
- integrates seamlessly with distributed hyperparameter optimization library Ray Tune
- comes with advanced fault tolerance handling mechanisms, and
- supports distributed dataframes and distributed data loading
All releases are tested on large clusters and workloads.
Installation
You can install the latest XGBoost-Ray release from PIP:
pip install xgboost_ray
If you'd like to install the latest master, use this command instead:
pip install git+https://github.com/ray-project/xgboost_ray.git#xgboost_ray
Usage
XGBoost-Ray provides a drop-in replacement for XGBoost's train function. To pass data, instead of using xgb.DMatrix you will have to use xgboost_ray.RayDMatrix.
Distributed training parameters are configured with a xgboost_ray.RayParams object. For instance, you can set the num_actors property to specify how many distributed actors you would like to use.
Here is a simplified example (which requires sklearn):
Training:
from xgboost_ray import RayDMatrix, RayParams, train
from sklearn.datasets import load_breast_cancer
train_x, train_y = load_breast_cancer(return_X_y=True)
train_set = RayDMatrix(train_x, train_y)
evals_result = {}
bst = train(
{
"objective": "binary:logistic",
"eval_metric": ["logloss", "error"],
},
train_set,
evals_result=evals_result,
evals=[(train_set, "train")],
verbose_eval=False,
ray_params=RayParams(
num_actors=2, # Number of remote actors
cpus_per_actor=1))
bst.save_model("model.xgb")
print("Final training error: {:.4f}".format(
evals_result["train"]["error"][-1]))
Prediction:
from xgboost_ray import RayDMatrix, RayParams, predict
from sklearn.datasets import load_breast_cancer
import xgboost as xgb
data, labels = load_breast_cancer(return_X_y=True)
dpred = RayDMatrix(data, labels)
bst = xgb.Booster(model_file="model.xgb")
pred_ray = predict(bst, dpred, ray_params=RayParams(num_actors=2))
print(pred_ray)
scikit-learn API
XGBoost-Ray also features a scikit-learn API fully mirroring pure XGBoost scikit-learn API, providing a completely drop-in replacement. The following estimators are available:
RayXGBClassifierRayXGRegressorRayXGBRFClassifierRayXGBRFRegressorRayXGBRanker
Example usage of RayXGBClassifier:
from xgboost_ray import RayXGBClassifier, RayParams
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
seed = 42
X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
X, y, train_size=0.25, random_state=42
)
clf = RayXGBClassifier(
n_jobs=4, # In XGBoost-Ray, n_jobs sets the number of actors
random_state=seed
)
# scikit-learn API will automatically conver the data
# to RayDMatrix format as needed.
# You can also pass X as a RayDMatrix, in which case
# y will be ignored.
clf.fit(X_train, y_train)
pred_ray = clf.predict(X_test)
print(pred_ray)
pred_proba_ray = clf.predict_proba(X_test)
print(pred_proba_ray)
# It is also possible to pass a RayParams object
# to fit/predict/predict_proba methods - will override
# n_jobs set during initialization
clf.fit(X_train, y_train, ray_params=RayParams(num_actors=2))
pred_ray = clf.predict(X_test, ray_params=RayParams(num_actors=2))
print(pred_ray)
Things to keep in mind:
n_jobsparameter controls the number of actors spawned. You can pass aRayParamsobject to thefit/predict/predict_probamethods as theray_paramsargument for greater control over resource allocation. Doing so will override the value ofn_jobswith the value ofray_params.num_actorsattribute. For more information, refer to the Resources section below.- By default
n_jobsis set to1, which means the training will not be distributed. Make sure to either setn_jobsto a higher value or pass aRayParamsobject as outlined above in order to take advantage of XGBoost-Ray's functionality. - After calling
fit, additional evaluation results (e.g. training time, number of rows, callback results) will be available underadditional_results_attribute. - XGBoost-Ray's scikit-learn API is based on XGBoost 1.4. While we try to support older XGBoost versions, please note that this library is only fully tested and supported for XGBoost >= 1.4.
For more information on the scikit-learn API, refer to the XGBoost documentation.
Data loading
Data is passed to XGBoost-Ray via a RayDMatrix object.
The RayDMatrix lazy loads data and stores it sharded in the Ray object store. The Ray XGBoost actors then access these shards to run their training on.
A RayDMatrix support various data and file types, like Pandas DataFrames, Numpy Arrays, CSV files and Parquet files.
Example loading multiple parquet files:
import glob
from xgboost_ray import RayDMatrix, RayFileType
# We can also pass a list of files
path = list(sorted(glob.glob("/data/nyc-taxi/*/*/*.parquet")))
# This argument will be passed to `pd.read_parquet()`
columns = [
"passenger_count",
"trip_distance", "pickup_longitude", "pickup_latitude",
"dropoff_longitude", "dropoff_latitude",
"fare_amount", "extra", "mta_tax", "tip_amount",
"tolls_amount", "total_amount"
]
dtrain = RayDMatrix(
path,
label="passenger_count", # Will select this column as the label
columns=columns,
filetype=RayFileType.PARQUET)
Hyperparameter Tuning
XGBoost-Ray integrates with Ray Tune to provide distributed hyperparameter tuning for your distributed XGBoost models. You can run multiple XGBoost-Ray training runs in parallel, each with a different hyperparameter configuration, and each training run parallelized by itself. All you have to do is move your training code to a function, and pass the function to tune.run. Internally, train will detect if tune is being used and will automatically report results to tune.
Example using XGBoost-Ray with Ray Tune:
from xgboost_ray import RayDMatrix, RayParams, train
from sklearn.datasets import load_breast_cancer
num_actors = 4
num_cpus_per_actor = 1
ray_params = RayParams(
num_actors=num_actors,
cpus_per_actor=num_cpus_per_actor)
def train_model(config):
train_x, train_y = load_breast_cancer(return_X_y=True)
train_set = RayDMatrix(train_x, train_y)
evals_result = {}
bst = train(
params=config,
dtrain=train_set,
evals_result=evals_result,
evals=[(train_set, "train")],
verbose_eval=False,
ray_params=ray_params)
bst.save_model("model.xgb")
from ray import tune
# Specify the hyperparameter search space.
config = {
"tree_method": "approx",
"objective": "binary:logistic",
"eval_metric": ["logloss", "error"],
"eta": tune.loguniform(1e-4, 1e-1),
"subsample": tune.uniform(0.5, 1.0),
"max_depth": tune.randint(1, 9)
}
# Make sure to use the `get_tune_resources` method to set the `resources_per_trial`
analysis = tune.run(
train_model,
config=config,
metric="train-error",
mode="min",
num_samples=4,
resources_per_trial=ray_params.get_tune_resources())
print("Best hyperparameters", analysis.best_config)
Also see examples/simple_tune.py for another example.
Fault tolerance
XGBoost-Ray leverages the stateful Ray actor model to enable fault tolerant training. There are currently two modes implemented.
Non-elastic training (warm restart)
When an actor or node dies, XGBoost-Ray will retain the state of the remaining actors. In non-elastic training, the failed actors will be replaced as soon as resources are available again. Only these actors will reload their parts of the data. Training will resume once all actors are ready for training again.
You can set this mode in the RayParams:
from xgboost_ray import RayParams
ray_params = RayParams(
elastic_training=False, # Use non-elastic training
max_actor_restarts=2, # How often are actors allowed to fail
)
Elastic training
In elastic training, XGBoost-Ray will continue training with fewer actors (and on fewer data) when a node or actor dies. The missing actors are staged in the background, and are reintegrated into training once they are back and loaded their data.
This mode will train on fewer data for a period of time, which can impact accuracy. In practice, we found these effects to be minor, especially for large shuffled datasets. The immediate benefit is that training time is reduced significantly to almost the same level as if no actors died. Thus, especially when data loading takes a large part of the total training time, this setting can dramatically speed up training times for large distributed jobs.
You can configure this mode in the RayParams:
from xgboost_ray import RayParams
ray_params = RayParams(
elastic_training=True, # Use elastic training
max_failed_actors=3, # Only allow at most 3 actors to die at the same time
max_actor_restarts=2, # How often are actors allowed to fail
)
Resources
By default, XGBoost-Ray tries to determine the number of CPUs available and distributes them evenly across actors.
In the case of very large clusters or clusters with many different machine sizes, it makes sense to limit the number of CPUs per actor by setting the cpus_per_actor argument. Consider always setting this explicitly.
The number of XGBoost actors always has to be set manually with the num_actors argument.
Multi GPU training
XGBoost-Ray enables multi GPU training. The XGBoost core backend will automatically leverage NCCL2 for cross-device communication. All you have to do is to start one actor per GPU.
For instance, if you have 2 machines with 4 GPUs each, you will want to start 8 remote actors, and set gpus_per_actor=1. There is usually no benefit in allocating less (e.g. 0.5) or more than one GPU per actor.
You should divide the CPUs evenly across actors per machine, so if your machines have 16 CPUs in addition to the 4 GPUs, each actor should have 4 CPUs to use.
from xgboost_ray import RayParams
ray_params = RayParams(
num_actors=8,
gpus_per_actor=1,
cpus_per_actor=4, # Divide evenly across actors per machine
)
How many remote actors should I use?
This depends on your workload and your cluster setup. Generally there is no inherent benefit of running more than one remote actor per node for CPU-only training. This is because XGBoost core can already leverage multiple CPUs via threading.
However, there are some cases when you should consider starting more than one actor per node:
- For multi GPU training, each GPU should have a separate remote actor. Thus, if your machine has 24 CPUs and 4 GPUs, you will want to start 4 remote actors with 6 CPUs and 1 GPU each
- In a heterogeneous cluster, you might want to find the greatest common divisor for the number of CPUs. E.g. for a cluster with three nodes of 4, 8, and 12 CPUs, respectively, you should set the number of actors to 6 and the CPUs per actor to 4.
Distributed data loading
XGBoost-Ray can leverage both centralized and distributed data loading.
In centralized data loading, the data is partitioned by the head node and stored in the object store. Each remote actor then retrieves their partitions by querying the Ray object store. Centralized loading is used when you pass centralized in-memory dataframes, such as Pandas dataframes or Numpy arrays, or when you pass a single source file, such as a single CSV or Parquet file.
from xgboost_ray import RayDMatrix
# This will use centralized data loading, as only one source file is specified
# `label_col` is a column in the CSV, used as the target label
ray_params = RayDMatrix("./source_file.csv", label="label_col")
In distributed data loading, each remote actor loads their data directly from the source (e.g. local hard disk, NFS, HDFS, S3), without a central bottleneck. The data is still stored in the object store, but locally to each actor. This mode is used automatically when loading data from multiple CSV or Parquet files. Please note that we do not check or enforce partition sizes in this case - it is your job to make sure the data is evenly distributed across the source files.
from xgboost_ray import RayDMatrix
# This will use distributed data loading, as four source files are specified
# Please note that you cannot schedule more than four actors in this case.
# `label_col` is a column in the Parquet files, used as the target label
ray_params = RayDMatrix([
"hdfs:///tmp/part1.parquet",
"hdfs:///tmp/part2.parquet",
"hdfs:///tmp/part3.parquet",
"hdfs:///tmp/part4.parquet",
], label="label_col")
Lastly, XGBoost-Ray supports distributed dataframe representations, such as Modin and Dask dataframes (used with Dask on Ray). Here, XGBoost-Ray will check on which nodes the distributed partitions are currently located, and will assign partitions to actors in order to minimize cross-node data transfer. Please note that we also assume here that partition sizes are uniform.
from xgboost_ray import RayDMatrix
# This will try to allocate the existing Modin partitions
# to co-located Ray actors. If this is not possible, data will
# be transferred across nodes
ray_params = RayDMatrix(existing_modin_df)
Data sources
| Type | Centralized loading | Distributed loading |
|---|---|---|
| Numpy array | Yes | No |
| Pandas dataframe | Yes | No |
| Single CSV | Yes | No |
| Multi CSV | Yes | Yes |
| Single Parquet | Yes | No |
| Multi Parquet | Yes | Yes |
| Petastorm | Yes | Yes |
| Ray MLDataset | Yes | Yes |
| Dask dataframe | Yes | Yes |
| Modin dataframe | Yes | Yes |
Memory usage
XGBoost uses a compute-optimized datastructure, the DMatrix, to hold training data. When converting a dataset to a DMatrix, XGBoost creates intermediate copies and ends up holding a complete copy of the full data. The data will be converted into the local dataformat (on a 64 bit system these are 64 bit floats.) Depending on the system and original dataset dtype, this matrix can thus occupy more memory than the original dataset.
The peak memory usage for CPU-based training is at least 3x the dataset size (assuming dtype float32 on a 64bit system) plus about 400,000 KiB for other resources, like operating system requirements and storing of intermediate results.
Example
- Machine type: AWS m5.xlarge (4 vCPUs, 16 GiB RAM)
- Usable RAM: ~15,350,000 KiB
- Dataset: 1,250,000 rows with 1024 features, dtype float32. Total size: 5,000,000 KiB
- XGBoost DMatrix size: ~10,000,000 KiB
This dataset will fit exactly on this node for training.
Note that the DMatrix size might be lower on a 32 bit system.
GPUs
Generally, the same memory requirements exist for GPU-based training. Additionally, the GPU must have enough memory to hold the dataset.
In the example above, the GPU must have at least 10,000,000 KiB (about 9.6 GiB) memory. However, empirically we found that using a DeviceQuantileDMatrix seems to show more peak GPU memory usage, possibly for intermediate storage when loading data (about 10%).
Best practices
In order to reduce peak memory usage, consider the following suggestions:
- Store data as
float32or less. More precision is often not needed, and keeping data in a smaller format will help reduce peak memory usage for initial data loading. - Pass the
dtypewhen loading data from CSV. Otherwise, floating point values will be loaded asnp.float64per default, increasing peak memory usage by 33%.
Placement Strategies
XGBoost-Ray leverages Ray's Placement Group API (https://docs.ray.io/en/master/placement-group.html) to implement placement strategies for better fault tolerance.
By default, a SPREAD strategy is used for training, which attempts to spread all of the training workers across the nodes in a cluster on a best-effort basis. This improves fault tolerance since it minimizes the number of worker failures when a node goes down, but comes at a cost of increased inter-node communication To disable this strategy, set the USE_SPREAD_STRATEGY environment variable to 0. If disabled, no particular placement strategy will be used.
Note that this strategy is used only when elastic_training is not used. If elastic_training is set to True, no placement strategy is used.
When XGBoost-Ray is used with Ray Tune for hyperparameter tuning, a PACK strategy is used. This strategy attempts to place all workers for each trial on the same node on a best-effort basis. This means that if a node goes down, it will be less likely to impact multiple trials.
When placement strategies are used, XGBoost-Ray will wait for 100 seconds for the required resources to become available, and will fail if the required resources cannot be reserved and the cluster cannot autoscale to increase the number of resources. You can change the PLACEMENT_GROUP_TIMEOUT_S environment variable to modify how long this timeout should be.
More examples
Fore complete end to end examples, please have a look at the examples folder:
- Simple sklearn breastcancer dataset example (requires
sklearn) - HIGGS classification example (download dataset (2.6 GB))
- HIGGS classification example with Parquet (uses the same dataset)
- Test data classification (uses a self-generated dataset)
![[RFC] Fallback for communication actor scheduling](https://avatars.githubusercontent.com/u/10364161?v=4)
997 Dec 30, 2022
2.5k Dec 28, 2022
27 Aug 19, 2022
338 Dec 26, 2022
519 Jan 3, 2023
517 Dec 31, 2022
2 Jan 18, 2022
1 Jan 19, 2022
164 Jan 4, 2023
2.5k Jan 6, 2023
2.3k Dec 30, 2022
393 Dec 27, 2022
14.5k Jan 7, 2023
12.9k Jan 7, 2023
3.3k Dec 28, 2022
23.6k Jan 3, 2023
1.6k Dec 29, 2022