You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@spark.apache.org by Keith Chapman <ke...@gmail.com> on 2018/02/22 08:13:15 UTC

Spark not releasing shuffle files in time (with very large heap)

Hi,

I'm benchmarking a spark application by running it for multiple iterations,
its a benchmark thats heavy on shuffle and I run it on a local machine with
a very large hear (~200GB). The system has a SSD. When running for 3 to 4
iterations I get into a situation that I run out of disk space on the /tmp
directory. On further investigation I was able to figure out that the
reason for this is that the shuffle files are still around, because I have
a very large hear GC has not happen and hence the shuffle files are not
deleted. I was able to confirm this by lowering the heap size and I see GC
kicking in more often and the size of /tmp stays under control. Is there
any way I could configure spark to handle this issue?

One option that I have is to have GC run more often by
setting spark.cleaner.periodicGC.interval to a much lower value. Is there a
cleaner solution?

Regards,
Keith.

http://keith-chapman.com

Re: Spark not releasing shuffle files in time (with very large heap)

Posted by Holden Karau <ho...@pigscanfly.ca>.
You can also look at the shuffle file cleanup tricks we do inside of the
ALS algorithm in Spark.

On Fri, Feb 23, 2018 at 6:20 PM, vijay.bvp <bv...@gmail.com> wrote:

> have you looked at
> http://apache-spark-user-list.1001560.n3.nabble.com/Limit-
> Spark-Shuffle-Disk-Usage-td23279.html
>
> and the post mentioned there
> https://forums.databricks.com/questions/277/how-do-i-avoid-
> the-no-space-left-on-device-error.html
>
> also try compressing the output
> https://spark.apache.org/docs/latest/configuration.html#
> compression-and-serialization
> spark.shuffle.compress
>
> thanks
> Vijay
>
>
>
> --
> Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/
>
> ---------------------------------------------------------------------
> To unsubscribe e-mail: user-unsubscribe@spark.apache.org
>
>


-- 
Twitter: https://twitter.com/holdenkarau

Re: Spark not releasing shuffle files in time (with very large heap)

Posted by "vijay.bvp" <bv...@gmail.com>.
have you looked at 
http://apache-spark-user-list.1001560.n3.nabble.com/Limit-Spark-Shuffle-Disk-Usage-td23279.html

and the post mentioned there
https://forums.databricks.com/questions/277/how-do-i-avoid-the-no-space-left-on-device-error.html

also try compressing the output
https://spark.apache.org/docs/latest/configuration.html#compression-and-serialization
spark.shuffle.compress

thanks
Vijay



--
Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org


Re: Spark not releasing shuffle files in time (with very large heap)

Posted by naresh Goud <na...@gmail.com>.
Got it. I understood issue in different way.



On Thu, Feb 22, 2018 at 9:19 PM Keith Chapman <ke...@gmail.com>
wrote:

> My issue is that there is not enough pressure on GC, hence GC is not
> kicking in fast enough to delete the shuffle files of previous iterations.
>
> Regards,
> Keith.
>
> http://keith-chapman.com
>
> On Thu, Feb 22, 2018 at 6:58 PM, naresh Goud <na...@gmail.com>
> wrote:
>
>> It would be very difficult to tell without knowing what is your
>> application code doing, what kind of transformation/actions performing.
>> From my previous experience tuning application code which avoids
>> unnecessary objects reduce pressure on GC.
>>
>>
>> On Thu, Feb 22, 2018 at 2:13 AM, Keith Chapman <ke...@gmail.com>
>> wrote:
>>
>>> Hi,
>>>
>>> I'm benchmarking a spark application by running it for multiple
>>> iterations, its a benchmark thats heavy on shuffle and I run it on a local
>>> machine with a very large hear (~200GB). The system has a SSD. When running
>>> for 3 to 4 iterations I get into a situation that I run out of disk space
>>> on the /tmp directory. On further investigation I was able to figure out
>>> that the reason for this is that the shuffle files are still around,
>>> because I have a very large hear GC has not happen and hence the shuffle
>>> files are not deleted. I was able to confirm this by lowering the heap size
>>> and I see GC kicking in more often and the size of /tmp stays under
>>> control. Is there any way I could configure spark to handle this issue?
>>>
>>> One option that I have is to have GC run more often by
>>> setting spark.cleaner.periodicGC.interval to a much lower value. Is there a
>>> cleaner solution?
>>>
>>> Regards,
>>> Keith.
>>>
>>> http://keith-chapman.com
>>>
>>
>>
>

Re: Spark not releasing shuffle files in time (with very large heap)

Posted by Keith Chapman <ke...@gmail.com>.
My issue is that there is not enough pressure on GC, hence GC is not
kicking in fast enough to delete the shuffle files of previous iterations.

Regards,
Keith.

http://keith-chapman.com

On Thu, Feb 22, 2018 at 6:58 PM, naresh Goud <na...@gmail.com>
wrote:

> It would be very difficult to tell without knowing what is your
> application code doing, what kind of transformation/actions performing.
> From my previous experience tuning application code which avoids
> unnecessary objects reduce pressure on GC.
>
>
> On Thu, Feb 22, 2018 at 2:13 AM, Keith Chapman <ke...@gmail.com>
> wrote:
>
>> Hi,
>>
>> I'm benchmarking a spark application by running it for multiple
>> iterations, its a benchmark thats heavy on shuffle and I run it on a local
>> machine with a very large hear (~200GB). The system has a SSD. When running
>> for 3 to 4 iterations I get into a situation that I run out of disk space
>> on the /tmp directory. On further investigation I was able to figure out
>> that the reason for this is that the shuffle files are still around,
>> because I have a very large hear GC has not happen and hence the shuffle
>> files are not deleted. I was able to confirm this by lowering the heap size
>> and I see GC kicking in more often and the size of /tmp stays under
>> control. Is there any way I could configure spark to handle this issue?
>>
>> One option that I have is to have GC run more often by
>> setting spark.cleaner.periodicGC.interval to a much lower value. Is
>> there a cleaner solution?
>>
>> Regards,
>> Keith.
>>
>> http://keith-chapman.com
>>
>
>

Re: Spark not releasing shuffle files in time (with very large heap)

Posted by naresh Goud <na...@gmail.com>.
It would be very difficult to tell without knowing what is your application
code doing, what kind of transformation/actions performing. From my
previous experience tuning application code which avoids unnecessary
objects reduce pressure on GC.


On Thu, Feb 22, 2018 at 2:13 AM, Keith Chapman <ke...@gmail.com>
wrote:

> Hi,
>
> I'm benchmarking a spark application by running it for multiple
> iterations, its a benchmark thats heavy on shuffle and I run it on a local
> machine with a very large hear (~200GB). The system has a SSD. When running
> for 3 to 4 iterations I get into a situation that I run out of disk space
> on the /tmp directory. On further investigation I was able to figure out
> that the reason for this is that the shuffle files are still around,
> because I have a very large hear GC has not happen and hence the shuffle
> files are not deleted. I was able to confirm this by lowering the heap size
> and I see GC kicking in more often and the size of /tmp stays under
> control. Is there any way I could configure spark to handle this issue?
>
> One option that I have is to have GC run more often by
> setting spark.cleaner.periodicGC.interval to a much lower value. Is there
> a cleaner solution?
>
> Regards,
> Keith.
>
> http://keith-chapman.com
>