[
https://issues.apache.org/jira/browse/MAPREDUCE-5649?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13843317#comment-13843317
]
Vitthal Gogate commented on MAPREDUCE-5649:
-------------------------------------------
I changed the code to make it long and it lowered the disk spills considerably
when enough memory is given to reducers. There are some other places where
Interger.MAX_VALUE is used which need be corrected.
In MergeManagerImpl.java following code is changed to use Long value,
- int maxInMemReduce = (int)Math.min(
- Runtime.getRuntime().maxMemory() * maxRedPer, Integer.MAX_VALUE);
+long maxInMemReduce = Math.min(
+ (long)(Runtime.getRuntime().maxMemory() * maxRedPer), Long.MAX_VALUE);
I will submit the patch soon.
> Reduce cannot use more than 2G memory for the final merge
> ----------------------------------------------------------
>
> Key: MAPREDUCE-5649
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-5649
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mrv2
> Affects Versions: trunk
> Reporter: stanley shi
>
> In the org.apache.hadoop.mapreduce.task.reduce.MergeManagerImpl.java file, in
> the finalMerge method:
> int maxInMemReduce = (int)Math.min(
> Runtime.getRuntime().maxMemory() * maxRedPer, Integer.MAX_VALUE);
>
> This means no matter how much memory user has, reducer will not retain more
> than 2G data in memory before the reduce phase starts.
--
This message was sent by Atlassian JIRA
(v6.1.4#6159)
