spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aurélien Bellet <aurelien.bel...@telecom-paristech.fr>
Subject Re: Memory-efficient successive calls to repartition()
Date Tue, 08 Sep 2015 21:21:39 GMT
What is strange is that if I remove the if condition (i.e., checkpoint 
at each iteration), then it basically works: non HDFS disk usage remains 
very small and stable throughout the execution.

If instead I checkpoint only every now and then (cf code in my previous 
email), then the disk usage grows regularly throughout the execution 
until no free space is available, despite the call to the GC.

Aurelien

Le 9/8/15 6:22 PM, Aurélien Bellet a écrit :
> Hi,
>
> This is what I tried:
>
> for i in range(1000):
>      print i
>      data2=data.repartition(50).cache()
>      if (i+1) % 10 == 0:
>          data2.checkpoint()
>      data2.first() # materialize rdd
>      data.unpersist() # unpersist previous version
>      sc._jvm.System.gc()
>      data=data2
>
> But unfortunately I do not get any significant improvement from the call
> to sc._jvm.System.gc()...
>
> I checked the WebUI and I have a single RDD in memory, so unpersist()
> works as expected but still no solution to trigger the cleaning of
> shuffle files...
>
> Aurélien
>
> Le 9/2/15 4:11 PM, alexis GILLAIN a écrit :
>> Just made some tests on my laptop.
>>
>> Deletion of the files is not immediate but a System.gc() call makes the
>> job on shuffle files of a checkpointed RDD.
>> It should solve your problem (`sc._jvm.System.gc()` in Python as pointed
>> in the databricks link in my previous message).
>>
>>
>> 2015-09-02 20:55 GMT+08:00 Aurélien Bellet
>> <aurelien.bellet@telecom-paristech.fr
>> <mailto:aurelien.bellet@telecom-paristech.fr>>:
>>
>>     Thanks a lot for the useful link and comments Alexis!
>>
>>     First of all, the problem occurs without doing anything else in the
>>     code (except of course loading my data from HDFS at the beginning) -
>>     so it definitely comes from the shuffling. You're right, in the
>>     current version, checkpoint files are not removed and take up some
>>     space in HDFS (this is easy to fix). But this is negligible compared
>>     to the non hdfs files which keeps growing as iterations go. So I
>>     agree with you that this must come from the shuffling operations: it
>>     seems that the shuffle files are not removed along the execution
>>     (they are only removed if I stop/kill the application), despite the
>>     use of checkpoint.
>>
>>     The class you mentioned is very interesting but I did not find a way
>>     to use it from pyspark. I will try to implement my own version,
>>     looking at the source code. But besides the queueing and removing of
>>     checkpoint files, I do not really see anything special there that
>>     could solve my issue.
>>
>>     I will continue to investigate this. Just found out I can use a
>>     command line browser to look at the webui (I cannot access the
>>     server in graphical display mode), this should help me understand
>>     what's going on. I will also try the workarounds mentioned in the
>>     link. Keep you posted.
>>
>>     Again, thanks a lot!
>>
>>     Best,
>>
>>     Aurelien
>>
>>
>>     Le 02/09/2015 14:15, alexis GILLAIN a écrit :
>>
>>         Aurélien,
>>
>>           From what you're saying, I can think of a couple of things
>>         considering
>>         I don't know what you are doing in the rest of the code :
>>
>>         - There is lot of non hdfs writes, it comes from the rest of
>>         your code
>>         and/or repartittion(). Repartition involve a shuffling and
>>         creation of
>>         files on disk. I would have said that the problem come from that
>>         but I
>>         just checked and checkpoint() is supposed to delete shuffle
>> files :
>>
>> https://forums.databricks.com/questions/277/how-do-i-avoid-the-no-space-left-on-device-error.html
>>
>>         (looks exactly as your problem so you could maybe try the others
>>         workarounds)
>>         Still, you may do a lot of shuffle in the rest of the code (you
>>         should
>>         see the amount of shuffle files written in the webui) and
>> consider
>>         increasing the disk space available...if you can do that.
>>
>>         - On the hdfs side, the class I pointed to has an update
>>         function which
>>         "automatically handles persisting and (optionally)
>> checkpointing, as
>>         well as unpersisting and removing checkpoint files". Not sure
>> your
>>         method for checkpointing remove previous checkpoint file.
>>
>>         In the end, does the disk space error come from hdfs growing or
>>         local
>>         disk growing ?
>>
>>         You should check the webui to identify which tasks spill data on
>>         disk
>>         and verify if the shuffle files are properly deleted when you
>>         checkpoint
>>         your rdd.
>>
>>
>>         Regards,
>>
>>
>>         2015-09-01 22:48 GMT+08:00 Aurélien Bellet
>>         <aurelien.bellet@telecom-paristech.fr
>>         <mailto:aurelien.bellet@telecom-paristech.fr>
>>         <mailto:aurelien.bellet@telecom-paristech.fr
>>         <mailto:aurelien.bellet@telecom-paristech.fr>>>:
>>
>>
>>              Dear Alexis,
>>
>>              Thanks again for your reply. After reading about
>>         checkpointing I
>>              have modified my sample code as follows:
>>
>>              for i in range(1000):
>>                   print i
>>                   data2=data.repartition(50).cache()
>>                   if (i+1) % 10 == 0:
>>                       data2.checkpoint()
>>                   data2.first() # materialize rdd
>>                   data.unpersist() # unpersist previous version
>>                   data=data2
>>
>>              The data is checkpointed every 10 iterations to a directory
>>         that I
>>              specified. While this seems to improve things a little bit,
>>         there is
>>              still a lot of writing on disk (appcache directory, shown
>>         as "non
>>              HDFS files" in Cloudera Manager) *besides* the checkpoint
>> files
>>              (which are regular HDFS files), and the application
>>         eventually runs
>>              out of disk space. The same is true even if I checkpoint at
>>         every
>>              iteration.
>>
>>              What am I doing wrong? Maybe some garbage collector setting?
>>
>>              Thanks a lot for the help,
>>
>>              Aurelien
>>
>>              Le 24/08/2015 10:39, alexis GILLAIN a écrit :
>>
>>                  Hi Aurelien,
>>
>>                  The first code should create a new RDD in memory at
>>         each iteration
>>                  (check the webui).
>>                  The second code will unpersist the RDD but that's not
>>         the main
>>                  problem.
>>
>>                  I think you have trouble due to long lineage as
>>         .cache() keep
>>                  track of
>>                  lineage for recovery.
>>                  You should have a look at checkpointing :
>>
>> https://github.com/JerryLead/SparkInternals/blob/master/markdown/english/6-CacheAndCheckpoint.md
>>
>>
>> https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/impl/PeriodicRDDCheckpointer.scala
>>
>>
>>                  You can also have a look at the code of others iterative
>>                  algorithms in
>>                  mlllib for best practices.
>>
>>                  2015-08-20 17:26 GMT+08:00 abellet
>>                  <aurelien.bellet@telecom-paristech.fr
>>         <mailto:aurelien.bellet@telecom-paristech.fr>
>>                  <mailto:aurelien.bellet@telecom-paristech.fr
>>         <mailto:aurelien.bellet@telecom-paristech.fr>>
>>                  <mailto:aurelien.bellet@telecom-paristech.fr
>>         <mailto:aurelien.bellet@telecom-paristech.fr>
>>
>>                  <mailto:aurelien.bellet@telecom-paristech.fr
>>         <mailto:aurelien.bellet@telecom-paristech.fr>>>>:
>>
>>                       Hello,
>>
>>                       For the need of my application, I need to
>> periodically
>>                  "shuffle" the
>>                       data
>>                       across nodes/partitions of a reasonably-large
>>         dataset. This
>>                  is an
>>                       expensive
>>                       operation but I only need to do it every now and
>> then.
>>                  However it
>>                       seems that
>>                       I am doing something wrong because as the
>>         iterations go the
>>                  memory usage
>>                       increases, causing the job to spill onto HDFS,
>> which
>>                  eventually gets
>>                       full. I
>>                       am also getting some "Lost executor" errors that I
>>         don't
>>                  get if I don't
>>                       repartition.
>>
>>                       Here's a basic piece of code which reproduces the
>>         problem:
>>
>>                       data =
>>
>>         sc.textFile("ImageNet_gist_train.txt",50).map(parseLine).cache()
>>                       data.count()
>>                       for i in range(1000):
>>                                data=data.repartition(50).persist()
>>                                # below several operations are done on
>> data
>>
>>
>>                       What am I doing wrong? I tried the following but
>>         it doesn't
>>                  solve
>>                       the issue:
>>
>>                       for i in range(1000):
>>                                data2=data.repartition(50).persist()
>>                                data2.count() # materialize rdd
>>                                data.unpersist() # unpersist previous
>> version
>>                                data=data2
>>
>>
>>                       Help and suggestions on this would be greatly
>>         appreciated!
>>                  Thanks a lot!
>>
>>
>>
>>
>>                       --
>>                       View this message in context:
>>
>> http://apache-spark-user-list.1001560.n3.nabble.com/Memory-efficient-successive-calls-to-repartition-tp24358.html
>>
>>                       Sent from the Apache Spark User List mailing list
>>         archive
>>                  at Nabble.com.
>>
>>
>>
>>
>> ---------------------------------------------------------------------
>>                       To unsubscribe, e-mail:
>>         user-unsubscribe@spark.apache.org
>>         <mailto:user-unsubscribe@spark.apache.org>
>>                  <mailto:user-unsubscribe@spark.apache.org
>>         <mailto:user-unsubscribe@spark.apache.org>>
>>                       <mailto:user-unsubscribe@spark.apache.org
>>         <mailto:user-unsubscribe@spark.apache.org>
>>                  <mailto:user-unsubscribe@spark.apache.org
>>         <mailto:user-unsubscribe@spark.apache.org>>>
>>                       For additional commands, e-mail:
>>         user-help@spark.apache.org <mailto:user-help@spark.apache.org>
>>                  <mailto:user-help@spark.apache.org
>>         <mailto:user-help@spark.apache.org>>
>>                       <mailto:user-help@spark.apache.org
>>         <mailto:user-help@spark.apache.org>
>>                  <mailto:user-help@spark.apache.org
>>         <mailto:user-help@spark.apache.org>>>
>>
>>
>>
>>
>>
>> ---------------------------------------------------------------------
>>              To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>         <mailto:user-unsubscribe@spark.apache.org>
>>              <mailto:user-unsubscribe@spark.apache.org
>>         <mailto:user-unsubscribe@spark.apache.org>>
>>              For additional commands, e-mail: user-help@spark.apache.org
>>         <mailto:user-help@spark.apache.org>
>>              <mailto:user-help@spark.apache.org
>>         <mailto:user-help@spark.apache.org>>
>>
>>
>>
>>     ---------------------------------------------------------------------
>>     To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>     <mailto:user-unsubscribe@spark.apache.org>
>>     For additional commands, e-mail: user-help@spark.apache.org
>>     <mailto:user-help@spark.apache.org>
>>
>>

---------------------------------------------------------------------
To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
For additional commands, e-mail: user-help@spark.apache.org


Mime
View raw message