[
https://issues.apache.org/jira/browse/SPARK-20797?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hyukjin Kwon updated SPARK-20797:
---------------------------------
Labels: bulk-closed (was: )
> mllib lda's LocalLDAModel's save: out of memory.
> -------------------------------------------------
>
> Key: SPARK-20797
> URL: https://issues.apache.org/jira/browse/SPARK-20797
> Project: Spark
> Issue Type: Bug
> Components: MLlib
> Affects Versions: 1.6.1, 1.6.3, 2.0.0, 2.0.2, 2.1.1
> Reporter: d0evi1
> Priority: Major
> Labels: bulk-closed
>
> when i try online lda model with large text data(nearly 1 billion chinese
> news' abstract), the training step went well, but the save step failed.
> something like below happened (etc. 1.6.1):
> problem 1.bigger than spark.kryoserializer.buffer.max. (turning bigger the
> param can fix problem 1, but next will lead problem 2),
> problem 2. exceed spark.akka.frameSize. (turning this param too bigger will
> fail for the reason out of memory, kill it, version > 2.0.0, exceeds max
> allowed: spark.rpc.message.maxSize).
> when topics num is large(set topic num k=200 is ok, but set k=300 failed),
> and vocab size is large(nearly 1000,000) too. this problem will appear.
> so i found word2vec's save function is similar to the LocalLDAModel's save
> function :
> word2vec's problem (use repartition(1) to save) has been fixed
> [https://github.com/apache/spark/pull/9989,], but LocalLDAModel still use:
> repartition(1). use single partition when save.
> word2vec's save method from latest code:
> https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/feature/Word2Vec.scala:
> val approxSize = (4L * vectorSize + 15) * numWords
> val nPartitions = ((approxSize / bufferSize) + 1).toInt
> val dataArray = model.toSeq.map { case (w, v) => Data(w, v) }
>
> spark.createDataFrame(dataArray).repartition(nPartitions).write.parquet(Loader.dataPath(path))
> but the code in mllib.clustering.LDAModel's LocalLDAModel's save:
> https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAModel.scala
> you'll see:
> val topicsDenseMatrix = topicsMatrix.asBreeze.toDenseMatrix
> val topics = Range(0, k).map { topicInd =>
> Data(Vectors.dense((topicsDenseMatrix(::, topicInd).toArray)),
> topicInd)
> }
>
> spark.createDataFrame(topics).repartition(1).write.parquet(Loader.dataPath(path))
> refer to word2vec's save (repartition(nPartitions)), i replace numWords to
> topic K, repartition(nPartitions) in the LocalLDAModel's save method,
> recompile the code, deploy the new lda's project with large data on our
> machine cluster, it works.
> hopes it will fixed in the next version.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
