spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mark Hamstra <m...@clearstorydata.com>
Subject Re: Spark scheduling mode
Date Fri, 02 Sep 2016 18:34:35 GMT
The comparator is used in `Pool#getSortedTaskSetQueue`.  The
`TaskSchedulerImpl` calls that on the rootPool when the TaskScheduler needs
to handle `resourceOffers` for available Executor cores.  Creation of the
`sortedTaskSets` is a recursive, nested sorting of the `Schedulable`
entities -- you can have pools within pools within pools within... if you
really want to, but they eventually bottom out in TaskSetManagers.  The
`sortedTaskSets` is a flattened queue of the TaskSets, and the available
cores are offered to those TaskSets in that queued order until the next
time the scheduler backend handles the available resource offers and a new
`sortedTaskSets` is generated.

On Fri, Sep 2, 2016 at 2:37 AM, enrico d'urso <E.DUrso@live.com> wrote:

> Thank you.
>
> May I know when that comparator is called?
> It looks like spark scheduler has not any form of preemption, am I right?
>
> Thank you
> ------------------------------
> *From:* Mark Hamstra <mark@clearstorydata.com>
> *Sent:* Thursday, September 1, 2016 8:44:10 PM
>
> *To:* enrico d'urso
> *Cc:* user@spark.apache.org
> *Subject:* Re: Spark scheduling mode
>
> Spark's FairSchedulingAlgorithm is not round robin: https://github.com/
> apache/spark/blob/master/core/src/main/scala/org/apache/spark/scheduler/
> SchedulingAlgorithm.scala#L43
>
> When at the scope of fair scheduling Jobs within a single Pool, the
> Schedulable entities being handled (s1 and s2) are TaskSetManagers, which
> are at the granularity of Stages, not Jobs.  Since weight is 1 and minShare
> is 0 for TaskSetManagers, the FairSchedulingAlgorithm for TaskSetManagers
> just boils down to prioritizing TaskSets (i.e. Stages) with the fewest
> number of runningTasks.
>
> On Thu, Sep 1, 2016 at 11:23 AM, enrico d'urso <E.DUrso@live.com> wrote:
>
>> I tried it before, but still I am not able to see a proper round robin
>> across the jobs I submit.
>> Given this:
>>
>> <pool name="production">
>>     <schedulingMode>FAIR</schedulingMode>
>>     <weight>1</weight>
>>     <minShare>2</minShare>
>>   </pool>
>>
>> Each jobs inside production pool should be scheduled in round robin way,
>> am I right?
>>
>> ------------------------------
>> *From:* Mark Hamstra <mark@clearstorydata.com>
>> *Sent:* Thursday, September 1, 2016 8:19:44 PM
>> *To:* enrico d'urso
>> *Cc:* user@spark.apache.org
>> *Subject:* Re: Spark scheduling mode
>>
>> The default pool (`<pool name = "default">`) can be configured like any
>> other pool: https://spark.apache.org/docs/latest/job-scheduling.
>> html#configuring-pool-properties
>>
>> On Thu, Sep 1, 2016 at 11:11 AM, enrico d'urso <E.DUrso@live.com> wrote:
>>
>>> Is there a way to force scheduling to be fair *inside* the default pool?
>>> I mean, round robin for the jobs that belong to the default pool.
>>>
>>> Cheers,
>>> ------------------------------
>>> *From:* Mark Hamstra <mark@clearstorydata.com>
>>> *Sent:* Thursday, September 1, 2016 7:24:54 PM
>>> *To:* enrico d'urso
>>> *Cc:* user@spark.apache.org
>>> *Subject:* Re: Spark scheduling mode
>>>
>>> Just because you've flipped spark.scheduler.mode to FAIR, that doesn't
>>> mean that Spark can magically configure and start multiple scheduling pools
>>> for you, nor can it know to which pools you want jobs assigned.  Without
>>> doing any setup of additional scheduling pools or assigning of jobs to
>>> pools, you're just dumping all of your jobs into the one available default
>>> pool (which is now being fair scheduled with an empty set of other pools)
>>> and the scheduling of jobs within that pool is still the default intra-pool
>>> scheduling, FIFO -- i.e., you've effectively accomplished nothing by only
>>> flipping spark.scheduler.mode to FAIR.
>>>
>>> On Thu, Sep 1, 2016 at 7:10 AM, enrico d'urso <E.DUrso@live.com> wrote:
>>>
>>>> I am building a Spark App, in which I submit several jobs (pyspark). I
>>>> am using threads to run them in parallel, and also I am setting:
>>>> conf.set("spark.scheduler.mode", "FAIR") Still, I see the jobs run
>>>> serially in FIFO way. Am I missing something?
>>>>
>>>> Cheers,
>>>>
>>>>
>>>> Enrico
>>>>
>>>
>>>
>>
>

Mime
View raw message