You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Joseph K. Bradley (JIRA)" <ji...@apache.org> on 2015/09/29 01:06:05 UTC
[jira] [Commented] (SPARK-10821) RandomForest serialization OOM
during findBestSplits
[ https://issues.apache.org/jira/browse/SPARK-10821?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14934298#comment-14934298 ]
Joseph K. Bradley commented on SPARK-10821:
-------------------------------------------
Hi, this is more a question for the user list, so I'll close it for now. But a few comments:
The real problem is that MLlib decision trees are meant for a relatively small number of features. They should work very well for a few thousand features, and could work for more but become fairly slow.
However, I'm working on a new implementation which should make training much faster with millions of features.
One suggestion: It sounds like your data are extremely sparse. I'd suggest hashing your feature vector to maybe 1000 features and try again.
> RandomForest serialization OOM during findBestSplits
> ----------------------------------------------------
>
> Key: SPARK-10821
> URL: https://issues.apache.org/jira/browse/SPARK-10821
> Project: Spark
> Issue Type: Bug
> Components: MLlib
> Affects Versions: 1.4.0, 1.5.0
> Environment: Amazon EC2 Linux
> Reporter: Jay Luan
> Labels: OOM, out-of-memory
>
> I am getting OOM during serialization for a relatively small dataset for a RandomForest. Even with spark.serializer.objectStreamReset at 1, It is still running out of memory when attempting to serialize my data.
> Stack Trace:
> Traceback (most recent call last):
> File "/root/random_forest/random_forest_spark.py", line 198, in <module>
> main()
> File "/root/random_forest/random_forest_spark.py", line 166, in main
> trainModel(dset)
> File "/root/random_forest/random_forest_spark.py", line 191, in trainModel
> impurity='gini', maxDepth=4, maxBins=32)
> File "/root/spark/python/lib/pyspark.zip/pyspark/mllib/tree.py", line 352, in trainClassifier
> File "/root/spark/python/lib/pyspark.zip/pyspark/mllib/tree.py", line 270, in _train
> File "/root/spark/python/lib/pyspark.zip/pyspark/mllib/common.py", line 130, in callMLlibFunc
> File "/root/spark/python/lib/pyspark.zip/pyspark/mllib/common.py", line 123, in callJavaFunc
> File "/root/spark/python/lib/py4j-0.8.2.1-src.zip/py4j/java_gateway.py", line 538, in __call__
> File "/root/spark/python/lib/py4j-0.8.2.1-src.zip/py4j/protocol.py", line 300, in get_return_value
> py4j.protocol.Py4JJavaError15/09/25 00:44:41 DEBUG BlockManagerSlaveEndpoint: Done removing RDD 7, response is 0
> 15/09/25 00:44:41 DEBUG BlockManagerSlaveEndpoint: Sent response: 0 to AkkaRpcEndpointRef(Actor[akka://sparkDriver/temp/$Mj])
> : An error occurred while calling o89.trainRandomForestModel.
> : java.lang.OutOfMemoryError
> at java.io.ByteArrayOutputStream.hugeCapacity(ByteArrayOutputStream.java:123)
> at java.io.ByteArrayOutputStream.grow(ByteArrayOutputStream.java:117)
> at java.io.ByteArrayOutputStream.ensureCapacity(ByteArrayOutputStream.java:93)
> at java.io.ByteArrayOutputStream.write(ByteArrayOutputStream.java:153)
> at java.io.ObjectOutputStream$BlockDataOutputStream.drain(ObjectOutputStream.java:1876)
> at java.io.ObjectOutputStream$BlockDataOutputStream.setBlockDataMode(ObjectOutputStream.java:1785)
> at java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1188)
> at java.io.ObjectOutputStream.writeObject(ObjectOutputStream.java:347)
> at org.apache.spark.serializer.JavaSerializationStream.writeObject(JavaSerializer.scala:44)
> at org.apache.spark.serializer.JavaSerializerInstance.serialize(JavaSerializer.scala:84)
> at org.apache.spark.util.ClosureCleaner$.ensureSerializable(ClosureCleaner.scala:301)
> at org.apache.spark.util.ClosureCleaner$.org$apache$spark$util$ClosureCleaner$$clean(ClosureCleaner.scala:294)
> at org.apache.spark.util.ClosureCleaner$.clean(ClosureCleaner.scala:122)
> at org.apache.spark.SparkContext.clean(SparkContext.scala:2021)
> at org.apache.spark.rdd.RDD$$anonfun$mapPartitions$1.apply(RDD.scala:703)
> at org.apache.spark.rdd.RDD$$anonfun$mapPartitions$1.apply(RDD.scala:702)
> at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:147)
> at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:108)
> at org.apache.spark.rdd.RDD.withScope(RDD.scala:306)
> at org.apache.spark.rdd.RDD.mapPartitions(RDD.scala:702)
> at org.apache.spark.mllib.tree.DecisionTree$.findBestSplits(DecisionTree.scala:625)
> at org.apache.spark.mllib.tree.RandomForest.run(RandomForest.scala:235)
> at org.apache.spark.mllib.tree.RandomForest$.trainClassifier(RandomForest.scala:291)
> at org.apache.spark.mllib.api.python.PythonMLLibAPI.trainRandomForestModel(PythonMLLibAPI.scala:742)
> at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
> at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:57)
> at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
> at java.lang.reflect.Method.invoke(Method.java:606)
> at py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:231)
> at py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:379)
> at py4j.Gateway.invoke(Gateway.java:259)
> at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:133)
> at py4j.commands.CallCommand.execute(CallCommand.java:79)
> at py4j.GatewayConnection.run(GatewayConnection.java:207)
> at java.lang.Thread.run(Thread.java:745)
> Details:
> My RDD is type MLLIB LabeledPoint objects, with each holding sparse vectors inside. This RDD has a total size of roughly 45MB. My sparse vector has a total length of ~15 million while only about 3000 or so are non-zeros. Works fine for up to sparse vector size 10 million.
> My cluster is setup on AWS such that my master is a r3.8xlarge along with two r3.4xlarge workers. Driver has ~190GB allocated to it while my RDD is ~45MB.
> Configurations as follows:
> spark version: 1.5.0
> -----------------------------------
> spark.executor.memory 32000m
> spark.driver.memory 230000m
> spark.driver.cores 10
> spark.executor.cores 5
> spark.executor.instances 17
> spark.driver.maxResultSize 0
> spark.storage.safetyFraction 1
> spark.storage.memoryFraction 0.9
> spark.storage.shuffleFraction 0.05
> spark.default.parallelism 128
> spark.serializer.objectStreamReset 1
> My original code is in python which I tried on 1.4.0 and 1.5.0, so I thought that maybe running something in scala may resolve the problem. I wrote a toy scala example and tested it on the same system yielding the same errors. Note the test code will most likely eventually throw an error due to the fact certain features are always 0 and MLLIB currently errors out during this operation.
> Running the following using spark-shell with my spark configuration gives me the OOM:
> --------------------------------------------------------------------------
> import scala.util.Random
> import scala.collection.mutable.ArrayBuffer
> import org.apache.spark.mllib.tree.RandomForest
> import org.apache.spark.mllib.tree.model.RandomForestModel
> import org.apache.spark.mllib.util.MLUtils
> import org.apache.spark.mllib.linalg.Vectors
> import org.apache.spark.mllib.regression.LabeledPoint
> val r = Random
> var size = 15000000
> var count = 3000
> val indptr = (1 to size by size/count).toArray
> val data = Seq.fill(count)(r.nextDouble()).toArray
> var dset = ArrayBuffer[LabeledPoint]()
> for (i <- 1 to 10) {
> dset += LabeledPoint(r.nextInt(2), Vectors.sparse(size, indptr, data));
> }
> val distData = sc.parallelize(dset)
> val splits = distData.randomSplit(Array(0.7, 0.3))
> val (trainingData, testData) = (splits(0), splits(1))
> // Train a RandomForest model.
> // Empty categoricalFeaturesInfo indicates all features are continuous.
> val numClasses = 2
> val categoricalFeaturesInfo = Map[Int, Int]()
> val numTrees = 3 // Use more in practice.
> val featureSubsetStrategy = "auto" // Let the algorithm choose.
> val impurity = "gini"
> val maxDepth = 4
> val maxBins = 32
> val model = RandomForest.trainClassifier(trainingData, numClasses, categoricalFeaturesInfo,
> numTrees, featureSubsetStrategy, impurity, maxDepth, maxBins)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org