Zhizhen Hou created MAPREDUCE-7080:
--------------------------------------
Summary: Default speculator won't sepculate the last several
submitted reduced task if the total task num is large
Key: MAPREDUCE-7080
URL: https://issues.apache.org/jira/browse/MAPREDUCE-7080
Project: Hadoop Map/Reduce
Issue Type: Improvement
Components: mrv2
Affects Versions: 2.7.5
Reporter: Zhizhen Hou
DefaultSpeculator speculates a task one time.
By default, the number of speculators is max(max(10, 0.01 * tasks.size), 0.1 *
running tasks)
I set mapreduce.job.reduce.slowstart.completedmaps = 1 to start reduce after
all the map tasks are finished.
The cluster has 1000 vcores, and the Job has 5000 reduce jobs.
At first, 1000 reduces tasks can run simultaneously, number of speculators can
speculator at most is 0.1 * 1000 = 100 tasks. Reduce tasks with less data can
over shortly, and speculator will speculator a task per second by default. The
task be speculated execution may be because the more data to be processed. It
will speculator 100 tasks within 100 seconds.
When 4900 reduces is over, If a reduce is executed with a lot of data be
processed and is put on a slow machine. The speculate opportunity is running
out, it will not be speculated. It can increase the execution time of job
significantly.
In short, it may waste the speculate opportunity at first only because the
execution time of reduce with less data to be processed as average time. At
end of job, there is no speculate opportunity available, especially last
several running tasks, judged the number of the running tasks .
In my opinion, the number of tasks be speculated can be judged by square of
finished task percent. Take an example, if ninety percent of the task is
finished, only 0.9*0.9 = 0.81 speculate opportunity can be used. It will leave
enough opportunity for latter tasks.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
