运行 Spark MLlib kmeans 时出现 OutOfMemory

Question

当我在大数据集上运行 Spark Kmeans 时，我总是遇到 OutOfMemory 错误。训练集大约 250GB，我有 10 个节点的 Spark 集群，每台机器有 16 个 CPU 和 150G 内存。我在每个节点上为作业分配了 100GB 内存，总共为 50 个 CPU。我将聚类中心设置为 100，迭代次数为 5。但是当代码在以下行上运行时，我得到了 OutOfMemory：

val model = KMeans.train(parsedData, numClusters, numIterations)

有没有我可以调整的参数来解决问题。

如果我设置较小的聚类中心数或迭代数就可以了。

我的代码如下：

val originalData = sc.textFile("hdfs://host/input.txt").cache()
val tupleData = originalData.map { x => (x.split(":")(0),x.split(":")(1)) }
val parsedData = tupleData.map { x => x._1 }.map(s => Vectors.dense(s.split(',').map(_.toDouble)))

val model = KMeans.train(parsedData, numClusters, numIterations, 1, initializationMode = KMeans.RANDOM)
val resultRdd = tupleData.map { p => (model.predict(Vectors.dense(p._1.split(',').map(_.toDouble))),p._2)}
resultRdd.sortByKey(true, 1).saveAsTextFile("hdfs://host/output.txt")

我的输入格式如下：

0.0,0.0,91.8,21.67,0.0 ... (the element number is 100K)
1.1,1.08,19.8,0.0,0.0 ... 
0.0,0.08,19.8,0.0,0.0 ...
...
The rows number is 600K.

我得到的异常如下：

scheduler.DAGScheduler: Submitting ShuffleMapStage 42 (MapPartitionsRDD[49] at map at KmeansTest.scala:47), which has no missing parents
Exception in thread "dag-scheduler-event-loop" java.lang.OutOfMemoryError: Java heap space
    at java.util.Arrays.copyOf(Arrays.java:2271)
    at java.io.ByteArrayOutputStream.grow(ByteArrayOutputStream.java:113)
    at java.io.ByteArrayOutputStream.ensureCapacity(ByteArrayOutputStream.java:93)
    at java.io.ByteArrayOutputStream.write(ByteArrayOutputStream.java:140)
    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)

Answer 1

By default Spark's Kmeans implementation use K_MEANS_PARALLEL initialization mode. Part of this mode run on the driver machine and can be extremely slow / cause OOM on the driver, depending on your data.

Try switching to RANDOM initialization mode.

val model = KMeans.train(parsedData, numClusters, numIterations, 1, initializationMode = KMeans.RANDOM)

Another thing to try is to increase your driver memory when you submit your application. For example, use the following command to set driver memory to 4G

spark-submit --conf "spark.driver.memory=4g" ...