[
https://issues.apache.org/jira/browse/SPARK-18548?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15691359#comment-15691359
]
Xiaoye Sun commented on SPARK-18548:
------------------------------------
Hi Yuhao,
Sorry for misleading you.
expElogbetaBc is the broadcast variable of expElogbeta.
It is unpersisted between the line of
val statsSum: BDM[Double] = stats.map(_._1).reduce(_ += _)
and
val gammat: BDM[Double] =
breeze.linalg.DenseMatrix.vertcat(stats.map(_._2).reduce(_ ++
_).map(_.toDenseMatrix): _*)
However, the later line also uses the expElogbetaBc which cause the worker
fetch expElogbetaBc again after it is unpersisted.
I suggest move the line of
expElogbetaBc.unpersist()
after the the line of
val gammat: BDM[Double] =
breeze.linalg.DenseMatrix.vertcat(stats.map(_._2).reduce(_ ++
_).map(_.toDenseMatrix): _*)
> OnlineLDAOptimizer reads the same broadcast data after deletion
> ---------------------------------------------------------------
>
> Key: SPARK-18548
> URL: https://issues.apache.org/jira/browse/SPARK-18548
> Project: Spark
> Issue Type: Improvement
> Components: MLlib
> Affects Versions: 1.6.1
> Reporter: Xiaoye Sun
> Priority: Trivial
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> In submitMiniBatch() called by OnlineLDAOptimizer, broadcast variable
> expElogbeta is deleted before its use in the second time, which causes the
> executor reads the same large broadcast data twice. I suggest to move the
> broadcast data deletion (expElogbetaBc.unpersist()) later.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
