[
https://issues.apache.org/jira/browse/HADOOP-4665?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12694663#action_12694663
]
Matei Zaharia commented on HADOOP-4665:
---------------------------------------
Hi Vivek,
A few comments:
* Is the 2-3 heartbeats for preemption really necessary? I imagine timeouts
will be on the order of minutes, so a few seconds won't make a big difference.
Although thinking of the timeout as an SLA is nice, it's also equally easy to
think of it as "this is when it can start killing tasks". To me, putting this
extra 2-3 heartbeats thing in seems like unnecessary complexity.
* The reason the preemption-enabled check is deep inside the method is to give
you the ability to turn preemption off but see the SHOULD_PREEMPT log messages
to figure out when your cluster *would* preempt tasks given certain settings.
We wanted this at Facebook so that we can add some timeouts, count the
SHOULD_PREEMPT messages over a week, and be sure that the settings chosen are
good without actually losing a lot of tasks if there's a mistake. I think this
is a good feature to keep for other people who are thinking of turning on
preemption.
* There actually is a way to set a default preemption timeout as in Joydeep's
comment - you can set defaultMinSharePreemptionTimeout in the XML file. The
code for this is in PoolManager.
* The default settings of preemptionEnabled=true and no timeouts are to make
preemption easy to configure gradually. We expect that most people will start
out not wanting preemption, because it creates an extra worry of "have we set
it too low". Then as people start running pools with "production" jobs (with
min share set), they may want to enable preemption just for these jobs. They
would be able to do that by just adding a preemptionTimeout entry to those
pools in the config, and it would be active without needing to restart the
JobTracker. Then if they see non-production jobs suffering, they could enable
the fairSharePreemptionTimeout, again without requiring a cluster restart. The
only reason to also provide a preemptionEnabled setting in the jobconf is for
the testing purpose I mentioned above, where an organization switching over to
preemption in production can figure out first whether it will kill too many
tasks. Overall, my goal with all the fair scheduler config is to make it as
easy as possible to use "out of the box". You don't need to define pools in
advance, you don't need to define min shares or weights in advance, you don't
need to decide when to use preemption in advance, etc, and the only setting you
need in mapred-site.xml is the one that sets Hadoop to use the fair scheduler.
Then as you decide you want the more advanced features, you enable them
gradually. I actually think there are strong advantages to this over your
proposal of having preemptionEnabled=false and having non-infinite default
timeouts so again I'd like more motivation before making such a large code
change. The other factor is that Facebook has been using the current version of
the preemption code and found the current features useful.
I'll take a look at your other comments later this week. Regarding the code
reuse in preemptTasks, it is indeed based on the one in the capacity scheduler,
but I'd like to make refactoring that a separate issue from this JIRA. The
right thing might be to have some of that functionality in TaskScheduler.
> Add preemption to the fair scheduler
> ------------------------------------
>
> Key: HADOOP-4665
> URL: https://issues.apache.org/jira/browse/HADOOP-4665
> Project: Hadoop Core
> Issue Type: New Feature
> Components: contrib/fair-share
> Reporter: Matei Zaharia
> Assignee: Matei Zaharia
> Fix For: 0.21.0
>
> Attachments: fs-preemption-v0.patch, hadoop-4665-v1.patch,
> hadoop-4665-v1b.patch, hadoop-4665-v2.patch, hadoop-4665-v3.patch,
> hadoop-4665-v4.patch
>
>
> Task preemption is necessary in a multi-user Hadoop cluster for two reasons:
> users might submit long-running tasks by mistake (e.g. an infinite loop in a
> map program), or tasks may be long due to having to process large amounts of
> data. The Fair Scheduler (HADOOP-3746) has a concept of guaranteed capacity
> for certain queues, as well as a goal of providing good performance for
> interactive jobs on average through fair sharing. Therefore, it will support
> preempting under two conditions:
> 1) A job isn't getting its _guaranteed_ share of the cluster for at least T1
> seconds.
> 2) A job is getting significantly less than its _fair_ share for T2 seconds
> (e.g. less than half its share).
> T1 will be chosen smaller than T2 (and will be configurable per queue) to
> meet guarantees quickly. T2 is meant as a last resort in case non-critical
> jobs in queues with no guaranteed capacity are being starved.
> When deciding which tasks to kill to make room for the job, we will use the
> following heuristics:
> - Look for tasks to kill only in jobs that have more than their fair share,
> ordering these by deficit (most overscheduled jobs first).
> - For maps: kill tasks that have run for the least amount of time (limiting
> wasted time).
> - For reduces: similar to maps, but give extra preference for reduces in the
> copy phase where there is not much map output per task (at Facebook, we have
> observed this to be the main time we need preemption - when a job has a long
> map phase and its reducers are mostly sitting idle and filling up slots).
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
