hadoop-common-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael Segel <michael_se...@hotmail.com>
Subject Re: Help me with architecture of a somewhat non-trivial mapreduce implementation
Date Fri, 20 Apr 2012 04:38:15 GMT
If the file is small enough you could read it in to a java object like a list and write your
own input format that takes a list object as its input and then lets you specify the number
of mappers.

On Apr 19, 2012, at 11:34 PM, Sky wrote:

> My file for the input to mapper is very small - as all it has is urls to list of manifests.
The task for mappers is to fetch each manifest, and then fetch files using urls from the manifests
and then process them.  Besides passing around lists of files, I am not really accessing the
disk. It should be RAM, network, and CPU (unzip, parsexml,extract attributes).
> 
> So is my only choice to break the input file and submit multiple files (if I have 15
cores, I should split the file with urls to 15 files? also how does it look in code?)? The
two drawbacks are - some cores might finish early and stay idle, and I don’t know how to
deal with dynamically increasing/decreasing cores.
> 
> Thx
> - Sky
> 
> -----Original Message----- From: Michael Segel
> Sent: Thursday, April 19, 2012 8:49 PM
> To: common-user@hadoop.apache.org
> Subject: Re: Help me with architecture of a somewhat non-trivial mapreduce implementation
> 
> How 'large' or rather in this case small is your file?
> 
> If you're on a default system, the block sizes are 64MB. So if your file ~<= 64MB,
you end up with 1 block, and you will only have 1 mapper.
> 
> 
> On Apr 19, 2012, at 10:10 PM, Sky wrote:
> 
>> Thanks for your reply.  After I sent my email, I found a fundamental defect - in
my understanding of how MR is distributed. I discovered that even though I was firing off
15 COREs, the map job - which is the most expensive part of my processing was run only on
1 core.
>> 
>> To start my map job, I was creating a single file with following data:
>> 1 storage:/root/1.manif.txt
>> 2 storage:/root/2.manif.txt
>> 3 storage:/root/3.manif.txt
>> ...
>> 4000 storage:/root/4000.manif.txt
>> 
>> I thought that each of the available COREs will be assigned a map job from top down
from the same file one at a time, and as soon as one CORE is done, it would get the next map
job. However, it looks like I need to split the file into the number of times. Now while that’s
clearly trivial to do, I am not sure how I can detect at runtime how many splits I need to
do, and also to deal with adding new CORES at runtime. Any suggestions? (it doesn't have to
be a file, it can be a list, etc).
>> 
>> This all would be much easier to debug, if somehow I could get my log4j logs for
my mappers and reducers. I can see log4j for my main launcher, but not sure how to enable
it for mappers and reducers.
>> 
>> Thx
>> - Akash
>> 
>> 
>> -----Original Message----- From: Robert Evans
>> Sent: Thursday, April 19, 2012 2:08 PM
>> To: common-user@hadoop.apache.org
>> Subject: Re: Help me with architecture of a somewhat non-trivial mapreduce implementation
>> 
>> From what I can see your implementation seems OK, especially from a performance perspective.
Depending on what storage: is it is likely to be your bottlekneck, not the hadoop computations.
>> 
>> Because you are writing files directly instead of relying on Hadoop to do it for
you, you may need to deal with error cases that Hadoop will normally hide from you, and you
will not be able to turn on speculative execution. Just be aware that a map or reduce task
may have problems in the middle, and be relaunched.  So when you are writing out your updated
manifest be careful to not replace the old one until the new one is completely ready and will
not fail, or you may lose data.  You may also need to be careful in your reduce if you are
writing directly to the file there too, but because it is not a read modify write, but just
a write it is not as critical.
>> 
>> --Bobby Evans
>> 
>> On 4/18/12 4:56 PM, "Sky USC" <skyusc@hotmail.com> wrote:
>> 
>> 
>> 
>> 
>> Please help me architect the design of my first significant MR task beyond "word
count". My program works well. but I am trying to optimize performance to maximize use of
available computing resources. I have 3 questions at the bottom.
>> 
>> Project description in an abstract sense (written in java):
>> * I have MM number of MANIFEST files available on storage:/root/1.manif.txt to 4000.manif.txt
>>   * Each MANIFEST in turn contains varilable number "EE" of URLs to EBOOKS (range
could be 10000 - 50,000 EBOOKS urls per MANIFEST) -- stored on storage:/root/1.manif/1223.folder/5443.Ebook.ebk
>> So we are talking about millions of ebooks
>> 
>> My task is to:
>> 1. Fetch each ebook, and obtain a set of 3 attributes per ebook (example: publisher,
year, ebook-version).
>> 2. Update each of the EBOOK entry record in the manifest - with the 3 attributes
(eg: ebook 1334 -> publisher=aaa year=bbb, ebook-version=2.01)
>> 3. Create a output file such that the named "<publisher>_<year>_<ebook-version>"
 contains a list of all "ebook urls" that met that criteria.
>> example:
>> File "storage:/root/summary/RANDOMHOUSE_1999_2.01.txt" contains:
>> storage:/root/1.manif/1223.folder/2143.Ebook.ebk
>> storage:/root/2.manif/2133.folder/5449.Ebook.ebk
>> storage:/root/2.manif/2133.folder/5450.Ebook.ebk
>> etc..
>> 
>> and File "storage:/root/summary/PENGUIN_2001_3.12.txt" contains:
>> storage:/root/19.manif/2223.folder/4343.Ebook.ebk
>> storage:/root/13.manif/9733.folder/2149.Ebook.ebk
>> storage:/root/21.manif/3233.folder/1110.Ebook.ebk
>> 
>> etc
>> 
>> 4. finally, I also want to output statistics such that:
>> <publisher>_<year>_<ebook-version>  <COUNT_OF_URLs>
>> PENGUIN_2001_3.12     250,111
>> RANDOMHOUSE_1999_2.01  11,322
>> etc
>> 
>> Here is how I implemented:
>> * My launcher gets list of MM manifests
>> * My Mapper gets one manifest.
>> --- It reads the manifest, within a WHILE loop,
>>  --- fetches each EBOOK,  and obtain attributes from each ebook,
>>  --- updates the manifest for that ebook
>>  --- context.write(new Text("RANDOMHOUSE_1999_2.01"), new Text("storage:/root/1.manif/1223.folder/2143.Ebook.ebk"))
>> --- Once all ebooks in the manifest are read, it saves the updated Manifest, and
exits
>> * My Reducer gets the "RANDOMHOUSE_1999_2.01" and a list of ebooks urls.
>> --- It writes a new file "storage:/root/summary/RANDOMHOUSE_1999_2.01.txt" with all
the storage urls for the ebooks
>> --- It also does a context.write(new Text("RANDOMHOUSE_1999_2.01"), new IntWritable(SUM_OF_ALL_EBOOK_URLS_FROM_THE_LIST))
>> 
>> As I mentioned, its working. I launch it on 15 elastic instances. I have three questions:
>> 1. Is this the best way to implement the MR logic?
>> 2. I dont know if each of the instances is getting one task or multiple tasks simultaneously
for the MAP portion. If it is not getting multiple MAP tasks, should I go with the route of
"multithreaded" reading of ebooks from each manifest? Its not efficient to read just one ebook
at a time per machine. Is "Context.write()" threadsafe?
>> 3. I can see log4j logs for main program, but no visibility into logs for Mapper
or Reducer. Any idea?
>> 
>> 
>> 
>> 
>> 
> 
> 


Mime
View raw message