[jira] [Commented] (SPARK-18890) Do all task serialization in CoarseGrainedExecutorBackend thread (rather than TaskSchedulerImpl)

["Imran Rashid (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-18890?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15805087#comment-15805087 ] 

Imran Rashid commented on SPARK-18890:
--------------------------------------

[~gq] I think you misunderstood my suggestion about using a broadcast for the task.  I'm not suggesting using a broadcast to contain *all* the task information, only the information which is shared across all tasks in a taskset. eg., the preferred location is ignored on the executor, so we wouldn't even bother serializing it either.  Conceptually, this means we'd have new classes  specially for sending the minimal necessary data to the executor, like:

{code}
/**
  * metadata about the taskset needed by the executor for all tasks in this taskset.  Subset of the
  * full data kept on the driver to make it faster to serialize and send to executors.
  */
class ExecutorTaskSetMeta(
  val stageId: Int,
  val stageAttemptId: Int,
  val properties: Properties,
  val addedFiles: Map[String, String],
  val addedJars: Map[String, String]
  // maybe task metrics here?
)

class ExecutorTaskData(
  val partitionId: Int,
  val attemptNumber: Int,
  val taskId: Long,
  val taskBinary: Broadcast[Array[Byte]],
  val taskSetMeta: Broadcast[ExecutorTaskSetMeta]
)
{code}

Then all the info you'd need to send to the executors would be a serialized version of ExecutorTaskData.  Furthermore, given the simplicity of that class, you could serialize manually, and then for each task you could just modify the first two ints & one long directly in the byte buffer.  (You could do the same trick for serialization even if ExecutorTaskSetMeta was not a broadcast, but that will keep the msgs small as well.)

There a bunch of details I'm skipping here: you'd also need to do some special handling for the TaskMetrics; the way tasks get started in the executor would change; you'd also need to refactor {{Task}} to let it get reconstructed from this information (or add more to ExecutorTaskSetMeta); and probably other details I'm overlooking now.

But if we really see task serialization as an issue, this seems like the right approach.

> Do all task serialization in CoarseGrainedExecutorBackend thread (rather than TaskSchedulerImpl)
> ------------------------------------------------------------------------------------------------
>
>                 Key: SPARK-18890
>                 URL: https://issues.apache.org/jira/browse/SPARK-18890
>             Project: Spark
>          Issue Type: Bug
>          Components: Scheduler
>    Affects Versions: 2.1.0
>            Reporter: Kay Ousterhout
>            Priority: Minor
>
>  As part of benchmarking this change: https://github.com/apache/spark/pull/15505 and alternatives, [~shivaram] and I found that moving task serialization from TaskSetManager (which happens as part of the TaskSchedulerImpl's thread) to CoarseGranedSchedulerBackend leads to approximately a 10% reduction in job runtime for a job that counted 10,000 partitions (that each had 1 int) using 20 machines.  Similar performance improvements were reported in the pull request linked above.  This would appear to be because the TaskSchedulerImpl thread is the bottleneck, so moving serialization to CGSB reduces runtime.  This change may *not* improve runtime (and could potentially worsen runtime) in scenarios where the CGSB thread is the bottleneck (e.g., if tasks are very large, so calling launch to send the tasks to the executor blocks on the network).
> One benefit of implementing this change is that it makes it easier to parallelize the serialization of tasks (different tasks could be serialized by different threads).  Another benefit is that all of the serialization occurs in the same place (currently, the Task is serialized in TaskSetManager, and the TaskDescription is serialized in CGSB).
> I'm not totally convinced we should fix this because it seems like there are better ways of reducing the serialization time (e.g., by re-using a single serialized object with the Task/jars/files and broadcasting it for each stage) but I wanted to open this JIRA to document the discussion.
> cc [~witgo]



--
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