[
https://issues.apache.org/jira/browse/MAPREDUCE-2083?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Faraz Ahmad updated MAPREDUCE-2083:
-----------------------------------
Description:
Shuffle delays can be large for mapreductions with lots of intermediate data.
Some of this shuffle delay can be overlapped with reduce if some of the reduce
computation is started on partial intermediate data received by a reduce.
Along these lines, the patch ??HADOOP-3226?? runs the combiner on the reduce
side to prune the data that goes to reduce. However, ??HADOOP-3226?? does not
achieve our goal of overlap with the shuffle because:
(1) In its original use of reducing intermediate data volume, the combiner
falls in the critical path at the map side. Therefore, the combiner is usually
a simple function which is too lightweight in its new use to achieve
sufficient overlap with the shuffle.
(2) Running the combiner at the reduce side is helpful in overlapping with the
shuffle only if the combiner's functionality is a major portion of the reduce
functionality -- otherwise running the combiner at the reduce side
achieves only modest overlap with the shuffle. In many mapreductions, the
combiner computation is often not part or only a small part of reduce
computation. Addressing both these points, reduces that are complex often
have heavier-weight computation than simple combining that can be overlapped
with the shuffle. This heavy-weight computation is specified by a user-supplied
"partial reduce" which performs the commutative/associative parts of reduce.
The idea is to run partial reduce on subsets of intermediate data as they
arrive at a reduce to overlap with the shuffle, and then run the full-blown
final reduce which re-reduces the partially-reduced data. Because the shuffle
delay is large for shuffle-heavy mapreductions, partial reduce that are
heavier-weight than simple combiner can be hidden
under the shuffle delay without extending the critical path of execution.
Finally, to further ensure that the partial reduce does not extend the critical
path, include two easily-tunable thresholds: One to start partial reduce only
after enough intermediate data has been received (e.g.
mapred.inmem.merge.threshold or a separately defined parameter) so that we do
not incur the overhead of invoking partial reduce on small data. Another
threshold to stop partial reduce after most of the intermediate data has been
received so that running partial reduce on the small remainder data
does not delay starting final reduce.
was:
Shuffle delays can be large for mapreductions with lots of intermediate data.
Some of this shuffle delay can be overlapped with reduce if some of the reduce
computation is started on partial intermediate data received by a reduce.
Along these lines, the patch ??HADOOP-3226?? runs the combiner
on the reduce side to prune the data that goes to reduce. However,
??HADOOP-3226?? does not
achieve our goal of overlap with the shuffle because: (1) In its original use
of reducing intermediate data volume,
the combiner falls in the critical path at the map side. Therefore, the
combiner is usually a simple function which is too lightweight in its new use
to achieve sufficient overlap with the shuffle.
(2) Running the combiner at the reduce side is helpful in overlapping with
the shuffle only if the combiner's functionality is a major portion of the
reduce functionality -- otherwise running the combiner at the reduce side
achieves only modest overlap with the shuffle.
In many mapreductions, the combiner computation is often not part or only a
small part of reduce computation.
Addressing both these points, reduces that are complex often
have heavier-weight computation than simple combining that can be overlapped
with the shuffle. This heavy-weight computation is specified by a
user-supplied "partial reduce" which performs the commutative/associative
parts of reduce. The idea is to run partial reduce on subsets of intermediate
data as they arrive at a reduce to overlap with the shuffle, and then run the
full-blown final reduce which re-reduces the partially-reduced data. Because
the shuffle delay is large for shuffle-heavy mapreductions,
partial reduce that are heavier-weight than simple combiner can be hidden
under the shuffle delay without extending the critical path of execution.
Finally, to further ensure that the partial reduce does not extend the
critical path, include two easily-tunable thresholds: One to start partial
reduce only after enough intermediate data has been received (e.g.
mapred.inmem.merge.threshold
or a separately defined parameter) so that we do not incur the overhead of
invoking partial reduce on small data.
Another threshold to stop partial reduce after most of the intermediate data
has been received so that running partial reduce on the small remainder data
does not delay starting final reduce.
> Run partial reduce instead of combiner at reduce node
> -----------------------------------------------------
>
> Key: MAPREDUCE-2083
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2083
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Reporter: Faraz Ahmad
> Fix For: 0.20.2
>
>
> Shuffle delays can be large for mapreductions with lots of intermediate data.
> Some of this shuffle delay can be overlapped with reduce if some of the
> reduce computation is started on partial intermediate data received by a
> reduce.
> Along these lines, the patch ??HADOOP-3226?? runs the combiner on the reduce
> side to prune the data that goes to reduce. However, ??HADOOP-3226?? does not
> achieve our goal of overlap with the shuffle because:
> (1) In its original use of reducing intermediate data volume, the combiner
> falls in the critical path at the map side. Therefore, the combiner is
> usually a simple function which is too lightweight in its new use to achieve
> sufficient overlap with the shuffle.
> (2) Running the combiner at the reduce side is helpful in overlapping with
> the shuffle only if the combiner's functionality is a major portion of the
> reduce functionality -- otherwise running the combiner at the reduce side
> achieves only modest overlap with the shuffle. In many mapreductions, the
> combiner computation is often not part or only a small part of reduce
> computation. Addressing both these points, reduces that are complex often
> have heavier-weight computation than simple combining that can be overlapped
> with the shuffle. This heavy-weight computation is specified by a
> user-supplied "partial reduce" which performs the commutative/associative
> parts of reduce. The idea is to run partial reduce on subsets of intermediate
> data as they arrive at a reduce to overlap with the shuffle, and then run
> the full-blown final reduce which re-reduces the partially-reduced data.
> Because the shuffle delay is large for shuffle-heavy mapreductions, partial
> reduce that are heavier-weight than simple combiner can be hidden
> under the shuffle delay without extending the critical path of execution.
> Finally, to further ensure that the partial reduce does not extend the
> critical path, include two easily-tunable thresholds: One to start partial
> reduce only after enough intermediate data has been received (e.g.
> mapred.inmem.merge.threshold or a separately defined parameter) so that we do
> not incur the overhead of invoking partial reduce on small data. Another
> threshold to stop partial reduce after most of the intermediate data has been
> received so that running partial reduce on the small remainder data
> does not delay starting final reduce.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
