[
https://issues.apache.org/jira/browse/TEZ-3209?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15390168#comment-15390168
]
Siddharth Seth commented on TEZ-3209:
-------------------------------------
bq. IMO, the main benefit of refactor is the "Async event handling" aspect and
thus the refactor seems worthy. What do you think?
Makes sense. I think the base class should be private to Tez for now though.
i.e. only plugins sitting within Tez should extend it.Would the plan be to
remove the global slow start rightnow, or later on when there's an actual need
to change the strategy. (A clear separation between scheduling for the vertex,
and event handling/routing may be available from Tez at a later point.)
> Support for fair custom data routing
> ------------------------------------
>
> Key: TEZ-3209
> URL: https://issues.apache.org/jira/browse/TEZ-3209
> Project: Apache Tez
> Issue Type: New Feature
> Reporter: Ming Ma
> Assignee: Ming Ma
> Attachments: TEZ-3209.patch, Tez-based demuxer for highly skewed
> category data.pdf
>
>
> This is based on offline discussion with [~gopalv], [~hitesh],
> [~jrottinghuis] and [~lohit] w.r.t. the support for efficient processing of
> highly skewed unordered partitioned mapper output. Our use case is to demux
> highly skewed unordered category data partitioned by category name. Gopal and
> Hitesh mentioned dynamically shuffled join scenario.
> One option we discussed is to leverage auto-parallelism feature with upfront
> over-partitioning. That means possible overhead to support large number
> partitions and unnecessary data movement as each reducer needs to get data
> from all mappers.
> Another alternative is to use custom {{DataMovementType}} which doesn't
> require each reducer to fetch data from all mappers. In that way, a large
> partition will be processed by several reducers, each of which will fetch
> data from a portion of mappers.
> For example, say there are 100 mappers each of which has 10 partitions (P1,
> ..., P10). Each mapper generates 100MB for its P10 and 1MB for each of its
> (P1, ... P9). The default SCATTER_GATHER routing means the reducer for P10
> has to process 10GB of input and becomes the bottleneck of the job. With the
> fair custom data routing, The P10 belonging to the first 10 mappers will be
> processed by one reducer with 1GB input data. The P10 belonging to the second
> 10 mappers will be processed by another reducer, etc.
> For further optimization, we can allocate the reducer on the same nodes as
> the mappers that it fetches data from.
> To support this, we need TEZ-3206 as well as customized data routing based on
> {{VertexManagerPlugin}} and {{EdgeManagerPluginOnDemand}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
