[
https://issues.apache.org/jira/browse/MAPREDUCE-6063?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
jiangyu updated MAPREDUCE-6063:
-------------------------------
Description:
In sortAndSpill of MapTask.java, size is calculated wrongly when bufend <
bufstart. we should change (bufvoid - bufend) + bufstart to (bufvoid -
bufstart) + bufend.
Should change
{code}
long size = (bufend >= bufstart
? bufend - bufstart
: (bufvoid - bufend) + bufstart) +
partitions * APPROX_HEADER_LENGTH;
{code}
to:
{code}
long size = (bufend >= bufstart
? bufend - bufstart
: (bufvoid - bufstart) + bufend) +
partitions * APPROX_HEADER_LENGTH;
{code}
It is because when wraparound happen (bufend < bufstart) , the size should
bufvoid - bufstart (bigger one) + bufend(small one).
You can find similar code implementation in MapTask.java:
{code}
mapOutputByteCounter.increment(valend >= keystart
? valend - keystart
: (bufvoid - keystart) + valend);
{code}
was:
In sortAndSpill of MapTask.java, size is calculated wrongly when bufend <
bufstart. we should change (bufvoid - bufend) + bufstart to (bufvoid -
bufstart) + bufend.
Should change
{code}
long size = (bufend >= bufstart
? bufend - bufstart
: (bufvoid - bufend) + bufstart) +
partitions * APPROX_HEADER_LENGTH;
{code}
to:
{code}
long size = (bufend >= bufstart
? bufend - bufstart
: (bufvoid - bufstart) + bufend) +
partitions * APPROX_HEADER_LENGTH;dd
{code}
It is because when wraparound happen (bufend < bufstart) , the size should
bufvoid - bufstart (bigger one) + bufend(small one).
You can find similar code implementation in MapTask.java:
{code}
mapOutputByteCounter.increment(valend >= keystart
? valend - keystart
: (bufvoid - keystart) + valend);
{code}
> In sortAndSpill of MapTask.java, size is calculated wrongly when bufend <
> bufstart.
> -----------------------------------------------------------------------------------
>
> Key: MAPREDUCE-6063
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-6063
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mrv1, mrv2
> Reporter: zhihai xu
> Assignee: zhihai xu
> Fix For: 2.6.0
>
> Attachments: MAPREDUCE-6063.000.patch, MAPREDUCE-6063.branch-1.patch
>
>
> In sortAndSpill of MapTask.java, size is calculated wrongly when bufend <
> bufstart. we should change (bufvoid - bufend) + bufstart to (bufvoid -
> bufstart) + bufend.
> Should change
> {code}
> long size = (bufend >= bufstart
> ? bufend - bufstart
> : (bufvoid - bufend) + bufstart) +
> partitions * APPROX_HEADER_LENGTH;
> {code}
> to:
> {code}
> long size = (bufend >= bufstart
> ? bufend - bufstart
> : (bufvoid - bufstart) + bufend) +
> partitions * APPROX_HEADER_LENGTH;
> {code}
> It is because when wraparound happen (bufend < bufstart) , the size should
> bufvoid - bufstart (bigger one) + bufend(small one).
> You can find similar code implementation in MapTask.java:
> {code}
> mapOutputByteCounter.increment(valend >= keystart
> ? valend - keystart
> : (bufvoid - keystart) + valend);
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
