-Re: Help me with architecture of a somewhat non-trivial mapreduce implementation
Robert Evans 2012-04-20, 15:16
You could also use the NLineInputFormat which will launch 1 mapper for every N (configurable) lines of input.
On 4/20/12 9:48 AM, "Sky" <[EMAIL PROTECTED]> wrote:
Thanks! That helped!
From: Michael Segel
Sent: Thursday, April 19, 2012 9:38 PM
To: [EMAIL PROTECTED]
Subject: Re: Help me with architecture of a somewhat non-trivial mapreduce
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.
> - Sky
> -----Original Message----- From: Michael Segel
> Sent: Thursday, April 19, 2012 8:49 PM
> To: [EMAIL PROTECTED]
> Subject: Re: Help me with architecture of a somewhat non-trivial mapreduce
> 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.
>> - Akash
>> -----Original Message----- From: Robert Evans
>> Sent: Thursday, April 19, 2012 2:08 PM
>> To: [EMAIL PROTECTED]
>> 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