hadoop-mapreduce-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Anirudh <techie.anir...@gmail.com>
Subject Re: instantiation of classes in MR
Date Mon, 09 Jan 2012 23:35:35 GMT
Harsh,

Appreciate you responding to the thread. Sorry got a bit held up with work,
hence couldnt reply.
I am not sure what do you mean by setup/cleanup procedures?
 In my response I actually meant latter i.e. API hooks 'setup/cleanup'. If
the APIs are called per split/partition, why do we need to do that? I
cannot think of a case where we would need to do the setup again for the
next split/partition, as it is the input data. As it is the setup/cleanup
cannot be dependent on the size of the partition/split. In JVM resuse
option multiple splits can be considered as a large single split. Something
like init()/service() API of servlets. Do you think there could be any
issues with this approach?


*Quotes: Harsh
That wouldn't be good isolation-wise. Resetting a lot of other relevant
parameters would be much more costly than reinitializing the whole task
(but not the JVM). I think https://issues.apache.org/jira/browse/HADOOP-249,
which introduced this, should help you*

This could be an answer to the question I asked earlier. As per me, reusing
the same instance of the Mapper would be good from optimization
perspective. Unless it is kind of mandated by the MR framework to run each
split/partition in isolation(even during JVM reuse). As I have yet not
looked at the code I will not be able to comment whether the
re-initializing that you are talking about could be avoided in case of JVM
reuse. Currently I am not sure what these parameters are that need to be
reset. I will surely look at the link that you provided and would want to
investigate further. Thanks for all the time.

Anirudh




On Mon, Jan 2, 2012 at 3:19 AM, Harsh J <harsh@cloudera.com> wrote:

> Hello Anirudh,
>
> On 02-Jan-2012, at 5:31 AM, Anirudh wrote:
>
> Any specific reason why setup is called for every task attempt. For
> optimization point of view, wouldnt it be good if the setup is called only
> once in case of JVM reuse.
>
>
> Note that the task setup/cleanup procedures are separate from the API
> hooks 'setup'/'cleanup'. The latter is guaranteed to the developer to be
> called per split/partition for pre-processing and post-processing execution.
>
> I have not yet looked at the implementation, in case of JVM reuse is the
> application Mapper instance reused or a new instance is created for every
> task attempt?
>
>
> That wouldn't be good isolation-wise. Resetting a lot of other relevant
> parameters would be much more costly than reinitializing the whole task
> (but not the JVM). I think
> https://issues.apache.org/jira/browse/HADOOP-249, which introduced this,
> should help you.
>
>
> My suggestion for Eyal would be to have a static field initializer
> expression in the Mapper to create the helper class instance. This will
> ensure that the helper class will be instantiated when the Mapper class is
> loaded.
>
>
> Yep this is possible, surely, and is a good advantage to using JVM reuse.
>
> On Sun, Jan 1, 2012 at 7:05 AM, Harsh J <harsh@cloudera.com> wrote:
>
>> You are guaranteed one setup call for every single task attempt. This
>> is regardless of JVM reuse being on or off. JVM reuse will cause no
>> issues with what Eyal is attempting to do.
>>
>> On Sun, Jan 1, 2012 at 5:49 PM, Anirudh <techie.anirudh@gmail.com> wrote:
>> > No problems Eyal.
>> >
>> > On  a second thought, for the JVM re-use the Mapper/Reducer instances
>> should
>> > be re-used, and the setup should be called only once. This makes sense
>> too
>> > as the JVM reuse is for the same job.
>> > You should be good with class instantiation even if the JVM reuse is
>> > enabled.
>> >
>> >
>> > On Sat, Dec 31, 2011 at 11:39 PM, Eyal Golan <egolan74@gmail.com>
>> wrote:
>> >>
>> >> Thank you very much for the detailed explanation Anirudh.
>> >>
>> >> I think that my question about node / VM was due to some lack of
>> knowledge
>> >> (I'm just starting to learn the Hadoop environment).
>> >> Regarding configuration of the nodes and clusters.
>> >> This is something that I am not doing by myself. We have a dedicated
>> team
>> >> for managing the Hadoop cluster and I'll ask them.
>> >>
>> >> I think that my question should have been: How many instances of the
>> >> 'helper' class will be created in a single VM.
>> >> And, as I understand, consider I am creating the helper in the setup /
>> >> configure method, there would be one.
>> >> And as long as it's stateless, I'm good.
>> >>
>> >> Thanks again,
>> >>
>> >> Eyal
>> >>
>> >>
>> >>
>> >> Eyal Golan
>> >> egolan74@gmail.com
>> >>
>> >> Visit: http://jvdrums.sourceforge.net/
>> >> LinkedIn: http://www.linkedin.com/in/egolan74
>> >> Skype: egolan74
>> >>
>> >> P  Save a tree. Please don't print this e-mail unless it's really
>> >> necessary
>> >>
>> >>
>> >>
>> >> On Sat, Dec 31, 2011 at 1:36 PM, Anirudh <techie.anirudh@gmail.com>
>> wrote:
>> >>>
>> >>> I just wanted to confirm where exactly you were planning to have the
>> >>> instantiation code, as it was not mentioned in your previous post. The
>> >>> location would have made difference. As you are doing it in the setup
>> of
>> >>> mapper/reducer, you are good.
>> >>>
>> >>> I was referring to the Task JVM Reuse option:
>> >>>
>> >>>
>> http://hadoop.apache.org/common/docs/current/mapred_tutorial.html#Task+JVM+Reuse
>> >>>
>> >>> It states that if the option to reuse JVM is enabled, the same Task
>> JVM
>> >>> will execute multiple tasks(i.e. map/reduce). I am not sure how this
>> is
>> >>> implemented, whether a new Mapper/Reducer is created for each task or
>> they
>> >>> too are re-reused.
>> >>> If a new instance is created each time, then the mapper/reducer and
>> all
>> >>> its reference will be marked for garbage collection and you would be
>> good.
>> >>> If the Mapper/Reducer instances are re-used then the setup should be
>> >>> called again creating another instance of your helper class.
>> >>>
>> >>> In my opinion the latter does not make sense, and the implementation
>> >>> would be according to the prior approach i.e. creation of a new
>> >>> Mapper/Reducer for each Task. But it would be interesting to check.
>> >>>
>> >>> As the classes in question are helper classes(stateless) you may not
>> get
>> >>> affected in terms of functionality.
>> >>>
>> >>> I am not clear on one of your statement:
>> >>>
>> >>> How many map tasks will be created? One per split or one per VM
>> (node)?
>> >>> Are you suggesting that although there would be one Mapper in the
>> node...
>> >>>
>> >>> Have you configured your node to have a single slot for map/reduce
>> task?
>> >>> If yes then there will be one Mapper/Reducer task in the node. If no
>> there
>> >>> could be more than one mapper/reducer in the node depending on lots
>> of other
>> >>> paramerters i.e. no of mappers/reducers slots allocated on the node,
>> no. of
>> >>> input splits etc. If the node is configured to run more than one
>> >>> Mapper/Reducer task the scheduler may choose to run more than one
>> task on
>> >>> the same node. The default is 2 Map & 2 Reduce tasks per node. And
>> for each
>> >>> task a new JVM is launched unless the JVM reuse option is enabled.
>> >>>
>> >>> Thanks,
>> >>> Anirudh
>> >>>
>> >>>
>> >>> On Sat, Dec 31, 2011 at 1:28 AM, Eyal Golan <egolan74@gmail.com>
>> wrote:
>> >>>>
>> >>>> My idea is to create that class in the setup / configure method
>> (depends
>> >>>> which Mapper / Reducer I will inherit from).
>> >>>>
>> >>>> I don't understand the 'reuse' option you are referring to.
>> >>>> How many map tasks will be created? One per split or one per VM
>> (node)?
>> >>>> Are you suggesting that although there would be one Mapper in the
>> node,
>> >>>> each new operator (or reflecting) will create a new instance?
>> >>>> Thus making lots of that instance?
>> >>>>
>> >>>> BTW,
>> >>>> these helper class I want to create are of course not going to be
>> >>>> stateful. They are defiantly 'helper' class that have some logic.
>> >>>>
>> >>>> Thanks,
>> >>>>
>> >>>> Eyal
>> >>>>
>> >>>> Eyal Golan
>> >>>> egolan74@gmail.com
>> >>>>
>> >>>> Visit: http://jvdrums.sourceforge.net/
>> >>>> LinkedIn: http://www.linkedin.com/in/egolan74
>> >>>> Skype: egolan74
>> >>>>
>> >>>> P  Save a tree. Please don't print this e-mail unless it's really
>> >>>> necessary
>> >>>>
>> >>>>
>> >>>>
>> >>>> On Sat, Dec 31, 2011 at 6:50 AM, Anirudh <techie.anirudh@gmail.com>
>> >>>> wrote:
>> >>>>>
>> >>>>> Where are you creating this new class. If it is in the map function,
>> >>>>> then it will be create a new object for each record in the split.
>> >>>>>
>> >>>>> Also you may need to see how the JVM reuse option works. I am
not
>> too
>> >>>>> sure of this and you may want to look at the code. If the option
>> for JVM
>> >>>>> reuse is set, then my understanding is for every task, a new
Map
>> task would
>> >>>>> be created and in that case the "new" operator will create another
>> instance
>> >>>>> even if this statement is not in the map function.
>> >>>>>
>> >>>>>
>> >>>>> On Fri, Dec 30, 2011 at 6:22 AM, Eyal Golan <egolan74@gmail.com>
>> wrote:
>> >>>>>>
>> >>>>>> Great News !!
>> >>>>>> Thanks for the info.
>> >>>>>>
>> >>>>>> So using reflection, I can inject different implementations
of
>> >>>>>> interfaces (services) for the mapper (or reducer).
>> >>>>>> And this way I can test a mapper (or reducer).
>> >>>>>> Just by reflecting a stub instead of a real implementation.
>> >>>>>>
>> >>>>>> Thanks,
>> >>>>>>
>> >>>>>>
>> >>>>>>
>> >>>>>> Eyal Golan
>> >>>>>> egolan74@gmail.com
>> >>>>>>
>> >>>>>> Visit: http://jvdrums.sourceforge.net/
>> >>>>>> LinkedIn: http://www.linkedin.com/in/egolan74
>> >>>>>> Skype: egolan74
>> >>>>>>
>> >>>>>> P  Save a tree. Please don't print this e-mail unless it's
really
>> >>>>>> necessary
>> >>>>>>
>> >>>>>>
>> >>>>>>
>> >>>>>> On Fri, Dec 30, 2011 at 2:50 PM, Harsh J <harsh@cloudera.com>
>> wrote:
>> >>>>>>>
>> >>>>>>> Eyal,
>> >>>>>>>
>> >>>>>>> Yes, it is right to think of each Task attempt being
one
>> individual
>> >>>>>>> JVM running individually on any added Node. Multiple
slots would
>> mean
>> >>>>>>> multiple VMs in parallel as well. Yes, your use of reflection
to
>> build your
>> >>>>>>> objects will work just fine -- its all user-side java
code that
>> is executed.
>> >>>>>>>
>> >>>>>>> On 30-Dec-2011, at 4:42 PM, Eyal Golan wrote:
>> >>>>>>>
>> >>>>>>> Hi,
>> >>>>>>>
>> >>>>>>> I want to understand a basic concept in MR.
>> >>>>>>>
>> >>>>>>> If a mapper creates an instance of some class (using
the 'new'
>> >>>>>>> operator), then the created class exists ONCE in the
VM of this
>> node.
>> >>>>>>> For each node.
>> >>>>>>> Correct?
>> >>>>>>>
>> >>>>>>> Now,
>> >>>>>>> what if instead of using the 'new' operator, the class
is created
>> >>>>>>> using reflection.
>> >>>>>>> Is it valid in a MR?
>> >>>>>>> Will only one instance of the created class be existing
in that
>> node?
>> >>>>>>>
>> >>>>>>> Thanks,
>> >>>>>>>
>> >>>>>>>
>> >>>>>>> Eyal
>> >>>>>>>
>> >>>>>>> Eyal Golan
>> >>>>>>> egolan74@gmail.com
>> >>>>>>>
>> >>>>>>> Visit: http://jvdrums.sourceforge.net/
>> >>>>>>> LinkedIn: http://www.linkedin.com/in/egolan74
>> >>>>>>> Skype: egolan74
>> >>>>>>>
>> >>>>>>> P  Save a tree. Please don't print this e-mail unless
it's really
>> >>>>>>> necessary
>> >>>>>>>
>> >>>>>>>
>> >>>>>>
>> >>>>>
>> >>>>
>> >>>
>> >>
>> >
>>
>>
>>
>> --
>> Harsh J
>>
>
>
>

Mime
View raw message