[
https://issues.apache.org/jira/browse/TEZ-3818?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ming Ma updated TEZ-3818:
-------------------------
Attachment: TEZ-3818.patch
Here is the draft patch. Besides the added unit test, I tested on a hadoop
cluster with real data.
> Support a new data routing policy for small partitions
> -------------------------------------------------------
>
> Key: TEZ-3818
> URL: https://issues.apache.org/jira/browse/TEZ-3818
> Project: Apache Tez
> Issue Type: Sub-task
> Reporter: Ming Ma
> Attachments: TEZ-3818.patch
>
>
> Under the existing fair shuffle manager data routing policies of
> fair_parallelism and increase_parallelism, small partitions (total size up to
> the max desirable limit) are processed together by a single destination task.
> We have the following use case that will prefer having one destination task
> process one small partition while still having multiple destination tasks
> process one large partition. When destination vertex is connected to
> MultiMROutput and the output format is parquet output format, each instance
> of parquet output stream consumes extra memory. So if a destination task ends
> up processing lots of small partitions, it ends up exceeding the task memory
> limit.
> With the new data routing policy, here is the summary of what each data
> routing policy does.
> * reduce_parallelism. The parallelism is decreased to a desired level by
> having one destination task process multiple consecutive partitions.
> * fair_parallelism. The parallelism is adjusted to a desired level by having
> one destination task process multiple consecutive small partitions and
> multiple destination tasks process one large partition.
> * The new increase_parallelism. The parallelism is increased to a desired
> level by having one destination task process each small partition and
> multiple destination tasks process one large partition.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
