[
https://issues.apache.org/jira/browse/HADOOP-657?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12598828#action_12598828
]
Ari Rabkin commented on HADOOP-657:
-----------------------------------
Here's my proposed fix:
1) Add a "free space on compute node" field to TaskTrackerStatus. This is
the real physical space available, plus the sum of (commitment - reservation)
for each running map task.
2) Add a "space used by this task" and "space reserved for task" to TaskStatus
as well.
3) Add a "space to reserve" to either Task or MapTask. This is computed by the
JobTracker, and used by the TaskTracker
4) Create a new ResourceConsumptionEstimator class, and have an instance of
that type for each JobInProgress. This will have, at a minimum,
reportCompletedMapTask(MapTaskStatus t) and estimateSpaceForMapTask(MapTask mt)
The implementation would probably be a thread that processes
asynchronously, and updates an atomic value that'll be either the estimated
space requirement, or else the estimated ratio between input size and output
size. Until sufficiently many maps have completed (10%, say) the size
estimate would just be the size of each map's input. Afterwards, we'll take
the 75th percentile of the measured blowup in task size.
5) Modify obtainNewMapTask to return null if the space available on the given
task tracker is less than the estimate of available space.
6) To avoid deadlocks if there are multiple jobs contending for space, abort
the job if too many trackers are rejected as having insufficient space.
-----
Thoughts?
> Free temporary space should be modelled better
> ----------------------------------------------
>
> Key: HADOOP-657
> URL: https://issues.apache.org/jira/browse/HADOOP-657
> Project: Hadoop Core
> Issue Type: Improvement
> Components: mapred
> Affects Versions: 0.7.2
> Reporter: Owen O'Malley
> Assignee: Ari Rabkin
>
> Currently, there is a configurable size that must be free for a task tracker
> to accept a new task. However, that isn't a very good model of what the task
> is likely to take. I'd like to propose:
> Map tasks: totalInputSize * conf.getFloat("map.output.growth.factor", 1.0) /
> numMaps
> Reduce tasks: totalInputSize * 2 * conf.getFloat("map.output.growth.factor",
> 1.0) / numReduces
> where totalInputSize is the size of all the maps inputs for the given job.
> To start a new task,
> newTaskAllocation + (sum over running tasks of (1.0 - done) * allocation)
> >=
> free disk * conf.getFloat("mapred.max.scratch.allocation", 0.90);
> So in English, we will model the expected sizes of tasks and only task tasks
> that should leave us a 10% margin. With:
> map.output.growth.factor -- the relative size of the transient data relative
> to the map inputs
> mapred.max.scratch.allocation -- the maximum amount of our disk we want to
> allocate to tasks.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
