hadoop-mapreduce-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Christoph Schmitz <christoph.schm...@1und1.de>
Subject RE: I need advice on whether my starting data needs to be in HDFS
Date Mon, 19 May 2014 14:57:18 GMT
Hi Steve,

I'll second David's opinion (if I get it correctly ;-) that importing your data first and
doing interesting processing later would be a good idea.

Of course you can read your data from NFS during your actual processing, as long as all of
your TaskTrackers have access to the NFS. Keep in mind, though, that importing your data in
a separate step gives you a lot more possibilities, e.g. to control the load on your NFS,
to do some simple preprocessing such as splitting the data by date, making sure your data
is ready for further processing (e.g. transformed into a splittable format such as SequenceFiles)
etc. In my team, we usually end up writing custom import code (rather that using, say, distcp),
but that pays off in the actual processing code.

IMHO, with 200 tasks (I assume that's what you meant) running in parallel you'll probably
saturate a "normal" NFS server, unless your computation is very expensive so that I/O is only
happening very sparsely in your CPU-bound task.

Regards,
Christoph

-----Original Message-----
From: David Rosenstrauch [mailto:darose@darose.net] 
Sent: Montag, 19. Mai 2014 16:33
To: mapreduce-user@hadoop.apache.org
Subject: Re: I need advice on whether my starting data needs to be in HDFS

That's pre-processing.  I.e., yes, before you run your job, when you 
push the file from local file system into HDFS then it is reading from a 
single machine.

However, presumably the processing your job does is the (far) more 
lengthy activity, and so running the job can take much less time if the 
data is in hdfs.

DR

On 05/19/2014 10:27 AM, Steve Lewis wrote:
> I understand that although I can write an InputFormat / splitter which
> starts with a file in NFS not HDFS.
> Also when I count import to hdfs as a part of processing haven't I gone to
> a single machine to read all data?
>
>
> On Mon, May 19, 2014 at 2:30 PM, David Rosenstrauch <darose@darose.net>wrote:
>
>> The reason why you want to copy to hdfs first is that hdfs splits the data
>> and distributes it across the nodes in the cluster.  So if your input data
>> is large, you'll get much better efficiency/speed in processing it if
>> you're processing it in a distributed manner.  (I.e., multiple machines
>> each processing a piece of it - multiple mappers.) I'd think that keeping
>> the in NFS would be quite slow.
>>
>> HTH,
>>
>> DR
>>
>>
>> On 05/15/2014 04:45 PM, Steve Lewis wrote:
>>
>>> I have a medium size data set in the terrabytes range that currently lives
>>> in the nfs file server of a medium institution. Every few months we want
>>> to
>>> run a chain of five Hadoop jobs on this data.
>>>      The cluster is medium sized - 40 nodes about 200 simultaneous jobs.
>>> The
>>> book says copy the data to HDFS and run the job. If I consider copy to
>>> hdfs
>>> and the first mapper as a single task I wonder if it is not as easy to
>>> have
>>> a custom reader reading from the NFS file system as a local file and skip
>>> the step of copying to hadoop.
>>>      While the read to the mapper may be slower, dropping the copy to hdfs
>>> could well make up the difference. Assume that after the job runs the data
>>> will be deleted from hdfs - the nfs system is the primary source and that
>>> cannot change. Also the job is not I/O limited - there is significant
>>> computation at each step
>>>
>>>       My questions are
>>>     1) are my assumptions correct and not copying the data may save time?
>>>     2) would 200 Hadoop jobs overwhelm a medium sized nfs system?
>>>
>>>
>>
>
>

Mime
View raw message