[jira] [Comment Edited] (TEZ-3222) Reduce messaging overhead for auto-reduce parallelism case

["Jonathan Eagles (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TEZ-3222?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15252781#comment-15252781 ] 

Jonathan Eagles edited comment on TEZ-3222 at 4/21/16 9:28 PM:
---------------------------------------------------------------

This patch addresses the above issue by reducing the numbers of messages by a factor of the reduction, keeping the number data movement events to be the same as before the reduction. It does this by create a new ranged data movement event, that contains fields that allow expansion on the task side as opposed to on the AM side.

In addition since the expansion is on the task side the duplicate processing of the payload and the empty partitions can be done once, allowing further optimization. For numbers sake, a 4k x 1k where the mappers generated all empty partitions and reduces auto-parallelized down to 1. (conf set slow start to 1.0 to measure actual reducer messaging cost). Before patch took 11 minutes to process 4M DME events and succeed 4M empty fetches. After patch, 6 seconds to expand the 4k ranged dmes and process the 4M empty expanded dmes.

\\cc [~rohini] [~jlowe] [~bikassaha] [~hitesh] [~sseth]


was (Author: jeagles):
This patch addresses the above issue by reducing the numbers of messages by a factor of the reduction, keeping the number data movement events to be the same as before the reduction. It does this by create a new ranged data movement event, that contains fields that allow expansion on the task side as opposed to on the AM side.

In addition since the expansion is on the task side the duplicate processing of the payload and the empty partitions can be done once, allowing further optimization. For numbers sake, a 4k x 1k where the mappers generated all empty partitions and reduces auto-parallelized down to 1. (conf set slow start to 1.0 to measure actual reducer messaging cost). Before patch took 11 minutes to process 4M DME events and succeed 4M empty fetches. After patch, 6 seconds to expand the 4k ranged dmes and process the 4M empty expanded dmes.

\\cc [~rohini] [~jlowe] []

> Reduce messaging overhead for auto-reduce parallelism case
> ----------------------------------------------------------
>
>                 Key: TEZ-3222
>                 URL: https://issues.apache.org/jira/browse/TEZ-3222
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: Jonathan Eagles
>            Assignee: Jonathan Eagles
>         Attachments: TEZ-3222.1.patch
>
>
> A dag with 15k x 1000k vertex may auto-reduce to 15k x 1. And while the data  size is appropriate for 1 task attempt, this results in an increase in task attempt message processing of 1000x.
> This jira aims to reduce the message processing in the auto-reduced task while keeping the amount of message processing in the AM the same or less.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)