hadoop-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From java8964 <java8...@hotmail.com>
Subject RE: AW: Extremely amount of memory and DB connections by MR Job
Date Mon, 29 Sep 2014 15:00:43 GMT
Here are my suggestions originally aims to improve the efficient:
1) In your case, you could use "StringBuilder", which has the append method, should be more
efficient to concatenate your string data in this case.2) What I mean to reuse the Text object
is as following:     public class mapper extends Mapper<> () {          private Text
data = new Text();          @Override           public void map(final Object key, final BasicDBObject
val, final Context context) throws IOException, InterruptedException {                   
    // instead of do "new Text(id)"                     // you can always use the following
way                     data.set(id);                     context.write(data, bsonWritable);
               }As you can see, you avoid to create lots, lots of Text object in the map method.
This method could be invoked a lot of times. In this way, you avoid asking GC to clean a lot
of Text object, by reusing the same Text object per map. I believe you can do the same for
BSONWritable. Check the javadoc for that class.3) 9G is a lot of heap for a map task. How
many map tasks your job generates? Are your source splitable? For one block data (I assume
it is 128M or 256M), I cannot image you need 9G heap for mapper. Your OOM maybe caused by
that your job runs out of physical memory of all the concurrent running mapper tasks.
1) How many total mapper tasks being generated in your job?2) How many data/task nodes you
have in your cluster? On the OOM node, how many mapper tasks being kicked off? You can find
all these information in the JobTracker in MR1, or AM in MR2.3) If each mapper assigned 9G
memory, and there are multi mappers running in the OOM node, how much real physical memory
you have? 4) You can see the input source for each mapper task in JobTracker or AM. If failed
mapper is always for the same block, then research that source data file. You need to have
real good reason to allocate 9G heap for a mapper task. Did you originally start from 1G?

Yong
From: Blanca.Hernandez@willhaben.at
To: user@hadoop.apache.org
Subject: AW: Extremely amount of memory and DB connections by MR Job
Date: Mon, 29 Sep 2014 14:16:24 +0000









Thanks for your answer.
To your questions:
 
1.      
When you claim  96G ram, I am not sure what do you mean?
It is not 96 Gb RAM, it is 9 Gb that our test server has available (is it too small?).
2.      
Your code is not efficient, as using the "+=" on String
I need (or at least I don´t have a better idea) the concatenation of strings for the emited
ID, since I want
 to group my objects by, e.g. Audi_A3_2010, another group Audi_A3_2011…. And so on. These
values are fields in the objects I get from the DB (BasicDBObject is a MongoDB class).
3.      
could have reused the Text object in your mapper
I am not sure if I understand your point. I create a new BSONWritable bsonWritable = new BSONWritable(val);
out of my
 data base object, since the one given by MongoDB is not mutable, hence not accepted by haddop
api as an outpu.
 
Now your other questions:
1) Are there any mappers successful?
Yes, but after a while, the job seems to need more memory, it runs very slow until it crashes.
2) The OOM mapper, is it always on the same block? If so, you need to dig into the source
data for that block, to think why it will cause OOM.
I am not sure about this. Is there a hint in the logs to figure it out?
3) Did you give reasonable heap size for the mapper? What it is?
9 Gb (too small??)
 
Best regards,

Blanca
 
 
 


Von: java8964 [mailto:java8964@hotmail.com]


Gesendet: Montag, 29. September 2014 15:43

An: user@hadoop.apache.org

Betreff: RE: Extremely amount of memory and DB connections by MR Job


 

I don't have any experience with MongoDB, but just gave my 2 cents here.

 


Your code is not efficient, as using the "+=" on String, and you could have reused the Text
object in your mapper, as it is a mutable class, to be reused and avoid creating it again
and again
 like "new Text()" in the mapper. My guess that BSONWritable should be a similar mutable class,
if it aims to be used like the rest Writable Hadoop class.


 


But even like that, it should just make your mapper run slower, as a lot of objects need to
be GC, instead of OOM.


 


When you claim  96G ram, I am not sure what do you mean? From what you said, it failed in
mapper stage, so let's focus on mapper. What max heap size you gave to the mapper task? I
don't think
 96G is the setting you mean to give to each mapper task. Otherwise, the only place I can
think is that there are millions of Strings to be appended in one record by "+=" and cause
the OOM.


 


You need to answer the following questions by yourself:


 


1) Are there any mappers successful?


2) The OOM mapper, is it always on the same block? If so, you need to dig into the source
data for that block, to think why it will cause OOM.


3) Did you give reasonable heap size for the mapper? What it is?


 


Yong


 




From:
Blanca.Hernandez@willhaben.at

To: user@hadoop.apache.org

Subject: Extremely amount of memory and DB connections by MR Job

Date: Mon, 29 Sep 2014 12:57:41 +0000

Hi, 
 
I am using a hadoop map reduce job + mongoDb.

It goes against a data base 252Gb big. During the job the amount of conexions is over 8000
and we gave already 9Gb RAM. The job is still crashing because of a OutOfMemory with
 only a 8% of the mapping done.
Are this numbers normal? Or did we miss something regarding configuration?
I attach my code, just in case the problem is with it.

 
Mapper:
 
public class AveragePriceMapper extends Mapper<Object, BasicDBObject, Text, BSONWritable>
{
    @Override
    public void map(final Object key, final BasicDBObject val, final Context context) throws
IOException, InterruptedException {
        String id = "";
        for(String propertyId : currentId.split(AveragePriceGlobal.SEPARATOR)){
            id += val.get(propertyId) + AveragePriceGlobal.SEPARATOR;
        }
        BSONWritable bsonWritable = new BSONWritable(val);
        context.write(new Text(id), bsonWritable);
    }
}
 
 
Reducer:
public class AveragePriceReducer extends Reducer<Text, BSONWritable, Text, Text>  {
    public void reduce(final Text pKey, final Iterable<BSONWritable> pValues, final
Context pContext) throws IOException, InterruptedException {
        while(pValues.iterator().hasNext() && continueLoop){
            BSONWritable next = pValues.iterator().next();
            //Make some calculations
        }        pContext.write(new Text(currentId), new Text(new MyClass(currentId, AveragePriceGlobal.COMMENT,
0, 0).toString()));
 
    }
}
 
The configuration includes a query which filters the number of objects to analyze (not the
252Gb will be analyzed).
 
Many thanks. Best regards,

Blanca




 		 	   		  
Mime
View raw message