Spark ML gradient boosted trees not using all nodes

问题

I'm using the Spark ML GBTClassifier in pyspark to train a binary classification model on a dataframe with ~400k rows and ~9k columns on an AWS EMR cluster. I'm comparing this against my current solution, which is running XGBoost on a huge EC2 that can fit the whole dataframe in memory.

My hope was that I could train (and score new observations) much faster in Spark because it would be distributed/parallel. However, when watch my cluster (through ganglia) I see that only 3-4 nodes have active CPU while the rest of the nodes are just sitting there. In fact, from the looks of it, it may be using only one node for the actual training.

I can't seem to find anything in the documentation about a node limit or partitions or anything that seems relevant to why this is happening. Maybe I'm just misunderstanding the implementation of the algorithm, but I assumed that it was implemented such a way that training could be parallelized to take advantage of the EMR/cluster aspect of Spark. If not, is there any advantage to doing it this way vs. just doing it in memory on a single EC2? I guess you don't have to load the data into memory, but that's not really much of an advantage.

Here is some boilerplate of my code. Thanks for any ideas!

import pyspark
from pyspark.sql import SparkSession
from pyspark.sql.functions import udf
from pyspark.sql.types import DoubleType
from pyspark.ml.classification import GBTClassifier
from pyspark.ml.evaluation import BinaryClassificationEvaluator

# Start Spark context:
sc = pyspark.SparkContext()
sqlContext = SparkSession.builder.enableHiveSupport().getOrCreate()

# load data
df = sqlContext.sql('SELECT label, features FROM full_table WHERE train = 1')
df.cache()
print("training data loaded: {} rows".format(df.count()))

test_df = sqlContext.sql('SELECT label, features FROM full_table WHERE train = 0')
test_df.cache()
print("test data loaded: {} rows".format(test_df.count()))


#Create evaluator
evaluator = BinaryClassificationEvaluator()
evaluator.setRawPredictionCol('prob')
evaluator.setLabelCol('label')

# train model
gbt = GBTClassifier(maxIter=100, 
                    maxDepth=3, 
                    stepSize=0.1,
                    labelCol="label", 
                    seed=42)

model = gbt.fit(df)


# get predictions
gbt_preds = model.transform(test_df)
gbt_preds.show(10)


# evaluate predictions
getprob=udf(lambda v:float(v[1]),DoubleType())
preds = gbt_preds.withColumn('prob', getprob('probability'))\
        .drop('features', 'rawPrediction', 'probability', 'prediction')
preds.show(10)

auc = evaluator.evaluate(preds)
auc

Sidenote: the tables I am using are already vectorized. The model runs with this code, it just runs slow (~10-15 mins to train) and only uses 3-4 (or maybe only one of the) cores.

回答1:

Thanks for the clarifying comments above.

It's not necessary that Spark's implementation is faster than XGBoost. In fact, I would expect what you're seeing.

The biggest factor is that XGBoost was designed and written specifically with Gradient Boosted Trees in mind. Spark, on the other hand is way more general purpose and most likely doesn't have the same kind of optimizations that XGBoost has. See here for a difference between XGBoost and scikit-learn's implementation of the classifier algorithm. If you want to really get into the details, you can read the paper and even the code behind XGBoost and Spark's implementations.

Remember, XGBoost is also parallel/distributed. It just uses multiple threads on the same machine. Spark helps you run the algorithm when the data doesn't fit on a single machine.

A couple other minor points I can think of are a) Spark does have a non-trivial startup time. Communication across different machines can also add up. b) XGBoost is written in C++ which is in general great for numerical computation.

As for why only 3-4 cores are being used by Spark, that depends on what your dataset size is, how it is being distributed across nodes, what is the number of executors that spark is launching, which stage is taking up most time, memory configuration, etc. You can use Spark UI to try and figure out what's going on. It's hard to say why that's happening for your dataset without looking at it.

Hope that helps.

Edit: I just found this great answer comparing execution times between a simple Spark application vs a standalone java application - https://stackoverflow.com/a/49241051/5509005. Same principles apply here as well, in fact much more so since XGBoost is highly optimized.

来源：https://stackoverflow.com/questions/49181179/spark-ml-gradient-boosted-trees-not-using-all-nodes