You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@spark.apache.org by Ryan Williams <ry...@gmail.com> on 2015/03/02 17:36:37 UTC
Re: bitten by spark.yarn.executor.memoryOverhead
For reference, the initial version of #3525
<https://github.com/apache/spark/pull/3525> (still open) made this fraction
a configurable value, but consensus went against that being desirable so I
removed it and marked SPARK-4665
<https://issues.apache.org/jira/browse/SPARK-4665> as "won't fix".
My team wasted a lot of time on this failure mode as well and has settled
in to passing "--conf spark.yarn.executor.memoryOverhead=1024" to most jobs
(that works out to 10-20% of --executor-memory, depending on the job).
I agree that learning about this the hard way is a weak part of the
Spark-on-YARN onboarding experience.
The fact that our instinct here is to increase the 0.07 minimum instead of
the alternate 384MB
<https://github.com/apache/spark/blob/3efd8bb6cf139ce094ff631c7a9c1eb93fdcd566/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala#L93>
minimum seems like evidence that the fraction is the thing we should allow
people to configure, instead of absolute amount that is currently
configurable.
Finally, do we feel confident that 0.1 is "always" enough?
On Sat, Feb 28, 2015 at 4:51 PM Corey Nolet <cj...@gmail.com> wrote:
> Thanks for taking this on Ted!
>
> On Sat, Feb 28, 2015 at 4:17 PM, Ted Yu <yu...@gmail.com> wrote:
>
>> I have created SPARK-6085 with pull request:
>> https://github.com/apache/spark/pull/4836
>>
>> Cheers
>>
>> On Sat, Feb 28, 2015 at 12:08 PM, Corey Nolet <cj...@gmail.com> wrote:
>>
>>> +1 to a better default as well.
>>>
>>> We were working find until we ran against a real dataset which was much
>>> larger than the test dataset we were using locally. It took me a couple
>>> days and digging through many logs to figure out this value was what was
>>> causing the problem.
>>>
>>> On Sat, Feb 28, 2015 at 11:38 AM, Ted Yu <yu...@gmail.com> wrote:
>>>
>>>> Having good out-of-box experience is desirable.
>>>>
>>>> +1 on increasing the default.
>>>>
>>>>
>>>> On Sat, Feb 28, 2015 at 8:27 AM, Sean Owen <so...@cloudera.com> wrote:
>>>>
>>>>> There was a recent discussion about whether to increase or indeed make
>>>>> configurable this kind of default fraction. I believe the suggestion
>>>>> there too was that 9-10% is a safer default.
>>>>>
>>>>> Advanced users can lower the resulting overhead value; it may still
>>>>> have to be increased in some cases, but a fatter default may make this
>>>>> kind of surprise less frequent.
>>>>>
>>>>> I'd support increasing the default; any other thoughts?
>>>>>
>>>>> On Sat, Feb 28, 2015 at 3:34 PM, Koert Kuipers <ko...@tresata.com>
>>>>> wrote:
>>>>> > hey,
>>>>> > running my first map-red like (meaning disk-to-disk, avoiding in
>>>>> memory
>>>>> > RDDs) computation in spark on yarn i immediately got bitten by a too
>>>>> low
>>>>> > spark.yarn.executor.memoryOverhead. however it took me about an hour
>>>>> to find
>>>>> > out this was the cause. at first i observed failing shuffles leading
>>>>> to
>>>>> > restarting of tasks, then i realized this was because executors
>>>>> could not be
>>>>> > reached, then i noticed in containers got shut down and reallocated
>>>>> in
>>>>> > resourcemanager logs (no mention of errors, it seemed the containers
>>>>> > finished their business and shut down successfully), and finally i
>>>>> found the
>>>>> > reason in nodemanager logs.
>>>>> >
>>>>> > i dont think this is a pleasent first experience. i realize
>>>>> > spark.yarn.executor.memoryOverhead needs to be set differently from
>>>>> > situation to situation. but shouldnt the default be a somewhat
>>>>> higher value
>>>>> > so that these errors are unlikely, and then the experts that are
>>>>> willing to
>>>>> > deal with these errors can tune it lower? so why not make the
>>>>> default 10%
>>>>> > instead of 7%? that gives something that works in most situations
>>>>> out of the
>>>>> > box (at the cost of being a little wasteful). it worked for me.
>>>>>
>>>>> ---------------------------------------------------------------------
>>>>> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>>>> For additional commands, e-mail: user-help@spark.apache.org
>>>>>
>>>>>
>>>>
>>>
>>
>
Re: bitten by spark.yarn.executor.memoryOverhead
Posted by Sean Owen <so...@cloudera.com>.
The problem is, you're left with two competing options then. You can
go through the process of deprecating the absolute one and removing it
eventually. You take away ability to set this value directly though,
meaning you'd have to set absolute values by depending on a % of what
you set your app memory too. I think there's non-trivial downside that
way too.
No value can always be right, or else it wouldn't be configurable. I
think of this one like any other param that's set in absolute terms,
but with an attempt to be smart about the default.
On Mon, Mar 2, 2015 at 4:36 PM, Ryan Williams
<ry...@gmail.com> wrote:
> For reference, the initial version of #3525 (still open) made this fraction
> a configurable value, but consensus went against that being desirable so I
> removed it and marked SPARK-4665 as "won't fix".
>
> My team wasted a lot of time on this failure mode as well and has settled in
> to passing "--conf spark.yarn.executor.memoryOverhead=1024" to most jobs
> (that works out to 10-20% of --executor-memory, depending on the job).
>
> I agree that learning about this the hard way is a weak part of the
> Spark-on-YARN onboarding experience.
>
> The fact that our instinct here is to increase the 0.07 minimum instead of
> the alternate 384MB minimum seems like evidence that the fraction is the
> thing we should allow people to configure, instead of absolute amount that
> is currently configurable.
>
> Finally, do we feel confident that 0.1 is "always" enough?
>
>
> On Sat, Feb 28, 2015 at 4:51 PM Corey Nolet <cj...@gmail.com> wrote:
>>
>> Thanks for taking this on Ted!
>>
>> On Sat, Feb 28, 2015 at 4:17 PM, Ted Yu <yu...@gmail.com> wrote:
>>>
>>> I have created SPARK-6085 with pull request:
>>> https://github.com/apache/spark/pull/4836
>>>
>>> Cheers
>>>
>>> On Sat, Feb 28, 2015 at 12:08 PM, Corey Nolet <cj...@gmail.com> wrote:
>>>>
>>>> +1 to a better default as well.
>>>>
>>>> We were working find until we ran against a real dataset which was much
>>>> larger than the test dataset we were using locally. It took me a couple days
>>>> and digging through many logs to figure out this value was what was causing
>>>> the problem.
>>>>
>>>> On Sat, Feb 28, 2015 at 11:38 AM, Ted Yu <yu...@gmail.com> wrote:
>>>>>
>>>>> Having good out-of-box experience is desirable.
>>>>>
>>>>> +1 on increasing the default.
>>>>>
>>>>>
>>>>> On Sat, Feb 28, 2015 at 8:27 AM, Sean Owen <so...@cloudera.com> wrote:
>>>>>>
>>>>>> There was a recent discussion about whether to increase or indeed make
>>>>>> configurable this kind of default fraction. I believe the suggestion
>>>>>> there too was that 9-10% is a safer default.
>>>>>>
>>>>>> Advanced users can lower the resulting overhead value; it may still
>>>>>> have to be increased in some cases, but a fatter default may make this
>>>>>> kind of surprise less frequent.
>>>>>>
>>>>>> I'd support increasing the default; any other thoughts?
>>>>>>
>>>>>> On Sat, Feb 28, 2015 at 3:34 PM, Koert Kuipers <ko...@tresata.com>
>>>>>> wrote:
>>>>>> > hey,
>>>>>> > running my first map-red like (meaning disk-to-disk, avoiding in
>>>>>> > memory
>>>>>> > RDDs) computation in spark on yarn i immediately got bitten by a too
>>>>>> > low
>>>>>> > spark.yarn.executor.memoryOverhead. however it took me about an hour
>>>>>> > to find
>>>>>> > out this was the cause. at first i observed failing shuffles leading
>>>>>> > to
>>>>>> > restarting of tasks, then i realized this was because executors
>>>>>> > could not be
>>>>>> > reached, then i noticed in containers got shut down and reallocated
>>>>>> > in
>>>>>> > resourcemanager logs (no mention of errors, it seemed the containers
>>>>>> > finished their business and shut down successfully), and finally i
>>>>>> > found the
>>>>>> > reason in nodemanager logs.
>>>>>> >
>>>>>> > i dont think this is a pleasent first experience. i realize
>>>>>> > spark.yarn.executor.memoryOverhead needs to be set differently from
>>>>>> > situation to situation. but shouldnt the default be a somewhat
>>>>>> > higher value
>>>>>> > so that these errors are unlikely, and then the experts that are
>>>>>> > willing to
>>>>>> > deal with these errors can tune it lower? so why not make the
>>>>>> > default 10%
>>>>>> > instead of 7%? that gives something that works in most situations
>>>>>> > out of the
>>>>>> > box (at the cost of being a little wasteful). it worked for me.
>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>>>>> For additional commands, e-mail: user-help@spark.apache.org
>>>>>>
>>>>>
>>>>
>>>
>>
>
---------------------------------------------------------------------
To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
For additional commands, e-mail: user-help@spark.apache.org
Re: bitten by spark.yarn.executor.memoryOverhead
Posted by Ted Yu <yu...@gmail.com>.
bq. that 0.1 is "always" enough?
The answer is: it depends (on use cases).
The value of 0.1 has been validated by several users. I think it is a
reasonable default.
Cheers
On Mon, Mar 2, 2015 at 8:36 AM, Ryan Williams <ryan.blake.williams@gmail.com
> wrote:
> For reference, the initial version of #3525
> <https://github.com/apache/spark/pull/3525> (still open) made this
> fraction a configurable value, but consensus went against that being
> desirable so I removed it and marked SPARK-4665
> <https://issues.apache.org/jira/browse/SPARK-4665> as "won't fix".
>
> My team wasted a lot of time on this failure mode as well and has settled
> in to passing "--conf spark.yarn.executor.memoryOverhead=1024" to most
> jobs (that works out to 10-20% of --executor-memory, depending on the job).
>
> I agree that learning about this the hard way is a weak part of the
> Spark-on-YARN onboarding experience.
>
> The fact that our instinct here is to increase the 0.07 minimum instead of
> the alternate 384MB
> <https://github.com/apache/spark/blob/3efd8bb6cf139ce094ff631c7a9c1eb93fdcd566/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala#L93>
> minimum seems like evidence that the fraction is the thing we should allow
> people to configure, instead of absolute amount that is currently
> configurable.
>
> Finally, do we feel confident that 0.1 is "always" enough?
>
>
> On Sat, Feb 28, 2015 at 4:51 PM Corey Nolet <cj...@gmail.com> wrote:
>
>> Thanks for taking this on Ted!
>>
>> On Sat, Feb 28, 2015 at 4:17 PM, Ted Yu <yu...@gmail.com> wrote:
>>
>>> I have created SPARK-6085 with pull request:
>>> https://github.com/apache/spark/pull/4836
>>>
>>> Cheers
>>>
>>> On Sat, Feb 28, 2015 at 12:08 PM, Corey Nolet <cj...@gmail.com> wrote:
>>>
>>>> +1 to a better default as well.
>>>>
>>>> We were working find until we ran against a real dataset which was much
>>>> larger than the test dataset we were using locally. It took me a couple
>>>> days and digging through many logs to figure out this value was what was
>>>> causing the problem.
>>>>
>>>> On Sat, Feb 28, 2015 at 11:38 AM, Ted Yu <yu...@gmail.com> wrote:
>>>>
>>>>> Having good out-of-box experience is desirable.
>>>>>
>>>>> +1 on increasing the default.
>>>>>
>>>>>
>>>>> On Sat, Feb 28, 2015 at 8:27 AM, Sean Owen <so...@cloudera.com> wrote:
>>>>>
>>>>>> There was a recent discussion about whether to increase or indeed make
>>>>>> configurable this kind of default fraction. I believe the suggestion
>>>>>> there too was that 9-10% is a safer default.
>>>>>>
>>>>>> Advanced users can lower the resulting overhead value; it may still
>>>>>> have to be increased in some cases, but a fatter default may make this
>>>>>> kind of surprise less frequent.
>>>>>>
>>>>>> I'd support increasing the default; any other thoughts?
>>>>>>
>>>>>> On Sat, Feb 28, 2015 at 3:34 PM, Koert Kuipers <ko...@tresata.com>
>>>>>> wrote:
>>>>>> > hey,
>>>>>> > running my first map-red like (meaning disk-to-disk, avoiding in
>>>>>> memory
>>>>>> > RDDs) computation in spark on yarn i immediately got bitten by a
>>>>>> too low
>>>>>> > spark.yarn.executor.memoryOverhead. however it took me about an
>>>>>> hour to find
>>>>>> > out this was the cause. at first i observed failing shuffles
>>>>>> leading to
>>>>>> > restarting of tasks, then i realized this was because executors
>>>>>> could not be
>>>>>> > reached, then i noticed in containers got shut down and reallocated
>>>>>> in
>>>>>> > resourcemanager logs (no mention of errors, it seemed the containers
>>>>>> > finished their business and shut down successfully), and finally i
>>>>>> found the
>>>>>> > reason in nodemanager logs.
>>>>>> >
>>>>>> > i dont think this is a pleasent first experience. i realize
>>>>>> > spark.yarn.executor.memoryOverhead needs to be set differently from
>>>>>> > situation to situation. but shouldnt the default be a somewhat
>>>>>> higher value
>>>>>> > so that these errors are unlikely, and then the experts that are
>>>>>> willing to
>>>>>> > deal with these errors can tune it lower? so why not make the
>>>>>> default 10%
>>>>>> > instead of 7%? that gives something that works in most situations
>>>>>> out of the
>>>>>> > box (at the cost of being a little wasteful). it worked for me.
>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>>>>> For additional commands, e-mail: user-help@spark.apache.org
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>