[Hadoop-common-user] HDFS vs. CIFS

CIFS is a file system, that doesn't scale particularly well, nor does it
support parallelisation of programs very well.

HDFS isn't quite a file system. At least not in the sense of something that
does all the things that you expect a file system to do (CRUD operations,
access control, meta-data). On the other hand, it does support one
particularly parallel execution model very nicely.

Not a valid comparison. CIFS is a remote file access protocol only. HDFS
is a file system (that comes bundled with a remote file access
protocol).

It may be possible to build a CIFS gateway for HDFS.

One interesting point of comparison at the protocol level is the level
of parallelism. Compared to HDFS protocol - CIFS exposes less
parallelism. DFS/CIFS has the concept of junction points that allows
directories from different storage servers to be stitched into one
namespace. There are commercial products that make this easy. However -
this allows parallelism at directory level only - whereas HDFS protocol
allows a single file to be distributed across different servers.

(And as was pointed out - CIFS supports many other file system
operations - ACLs, oplocks and what not that HDFS doesn't).

-----Original Message-----
From: TREVORSTEWART@UP.COM
Sent: Monday, October 15, 2007 12:24 PM
To: hadoop-user@lucene.apache.org
Subject: HDFS vs. CIFS


I would like someone to compare and contrast CIFS and HDFS? Or...if
that
is not a valid comparison...please explain to me why it's not a valid
comparison.

Thanks,
Trevor

.
This message and any attachments contain information from Union Pacific
which may be confidential and/or privileged.
If you are not the intended recipient, be aware that any disclosure,
copying, distribution or use of the contents of this message is strictly
prohibited by law. If you receive this message in error, please contact
the sender immediately and delete the message and any attachments.

Hmmm...OK...

Let me explain my requirements here and see if you all can tell me if
Hadoop provides the functionality I need.

I'm building a highly perfomant, highly available (no less than 4 9's), raw
storage subsystem. It will be write once for the initial dataset (binary
data) but will have the ability to maintain metadata associated to the
binary data. The metadata will be "queryiable" and therefore indexed
(want to use Lucene for this purpose). It must have the ability to store
petrabytes of data. We will use either NetApps or DMX3 storage media.

Please discuss...






"Joydeep Sen
Sarma"
<jssarma@facebook To
.com> <hadoop-user@lucene.apache.org>
cc
10/15/2007 05:20
PM Subject
RE: HDFS vs. CIFS

Please respond to
hadoop-user@lucen
e.apache.org






Not a valid comparison. CIFS is a remote file access protocol only. HDFS
is a file system (that comes bundled with a remote file access
protocol).

It may be possible to build a CIFS gateway for HDFS.

One interesting point of comparison at the protocol level is the level
of parallelism. Compared to HDFS protocol - CIFS exposes less
parallelism. DFS/CIFS has the concept of junction points that allows
directories from different storage servers to be stitched into one
namespace. There are commercial products that make this easy. However -
this allows parallelism at directory level only - whereas HDFS protocol
allows a single file to be distributed across different servers.

(And as was pointed out - CIFS supports many other file system
operations - ACLs, oplocks and what not that HDFS doesn't).

-----Original Message-----
From: TREVORSTEWART@UP.COM
Sent: Monday, October 15, 2007 12:24 PM
To: hadoop-user@lucene.apache.org
Subject: HDFS vs. CIFS


I would like someone to compare and contrast CIFS and HDFS? Or...if
that
is not a valid comparison...please explain to me why it's not a valid
comparison.

Thanks,
Trevor

.
This message and any attachments contain information from Union Pacific
which may be confidential and/or privileged.
If you are not the intended recipient, be aware that any disclosure,
copying, distribution or use of the contents of this message is strictly
prohibited by law. If you receive this message in error, please contact
the sender immediately and delete the message and any attachments.




. This message and any attachments contain information from Union Pacific which may be confidential and/or privileged.
If you are not the intended recipient, be aware that any disclosure, copying, distribution or use of the contents of this message is strictly prohibited by law. If you receive this message in error, please contact the sender immediately and delete the message and any attachments.

First, it is PETAbytes, not petRabytes.

Secondly, if you are committed to using NetApps or DMX3, then you really
don't need (or want HDFS).

Thirdly, if you are committed to using a distributed file store like HDFS
(or MogileFS or KFS), then you don't need NetApps. Distributed file systems
were designed exactly to eliminate the need for highly engineered storage
systems by allowing the use of entire redundant computers rather than
cleverly interconnected disks.

So you really have two classes of designs:

A) traditional big iron

B) trendy, but not entirely ready for prime time distributed file stores
like HDFS

The first option will probably work and will cost about 2x more (based on my
experience, your mileage will vary). The second option will require more
hand-holding and won't come with a support contract, but you would be able
to do some things with it that are impossible in a traditional sense.


My guess is that if you are still asking basic questions like this that are
answered in the FAQ, then you will be better off paying NetApp for
engineering time than building this system on your own.

Well then...color me humbled Mr. Dunning.

I apologize for monopolizing your quite obviously precious time.

BTW...I don't believe these questions are answered in the FAQ.

Thank you for making the open source experience SO enjoyable.





Ted Dunning
<tdunning@veoh.co
m> To
<hadoop-user@lucene.apache.org>
10/16/2007 12:32 cc
PM
Subject
Re: HDFS vs. CIFS
Please respond to
hadoop-user@lucen
e.apache.org








First, it is PETAbytes, not petRabytes.

Secondly, if you are committed to using NetApps or DMX3, then you really
don't need (or want HDFS).

Thirdly, if you are committed to using a distributed file store like HDFS
(or MogileFS or KFS), then you don't need NetApps. Distributed file
systems
were designed exactly to eliminate the need for highly engineered storage
systems by allowing the use of entire redundant computers rather than
cleverly interconnected disks.

So you really have two classes of designs:

A) traditional big iron

B) trendy, but not entirely ready for prime time distributed file stores
like HDFS

The first option will probably work and will cost about 2x more (based on
my
experience, your mileage will vary). The second option will require more
hand-holding and won't come with a support contract, but you would be able
to do some things with it that are impossible in a traditional sense.


My guess is that if you are still asking basic questions like this that are
answered in the FAQ, then you will be better off paying NetApp for
engineering time than building this system on your own.




. This message and any attachments contain information from Union Pacific which may be confidential and/or privileged.
If you are not the intended recipient, be aware that any disclosure, copying, distribution or use of the contents of this message is strictly prohibited by law. If you receive this message in error, please contact the sender immediately and delete the message and any attachments.

Apologies off-list. That wasn't intended to be rude.

http://lucene.apache.org/hadoop/hdfs_design.html
It's not the FAQ, but this document should provide a pretty good
description of the (intended) features of HDFS. It does not provide a
complete contrast against other distributed file systems, but it gives
you an idea of what it was designed for, and how and why the interface
differs from normal POSIX filesystems. Based on the nature of your
application, this should help you make a decision on whether HDFS is
right for you.
http://labs.google.com/papers/gfs.html
The Google GFS paper also provides a good idea of how this type of
filesystem differs from what you are used to.
Best,
Garth

Fair enough...

I appreciate your reply and apologize for my misinterpretation of your
intent. Someone else pointed me to the architecture documentation which I
shall peruse inorder to gain a better understanding of this product.

The primary feature that attracted me to Hadoop was the ability to maintain
a single namespace across resources. This will become increasingly
important as we add logical volumes to our storage array, whether they be
NetApps, DMX3, or commodity hardware (servers). I have, up until I came
across hadoop, been focusing primarily on CIFS and want to further
investigate other distributed file systems in order to either rule them out
or to further realize their capabilities and how they may apply to the
problem at hand.

Thank you all for your replies.

Trevor Stewart
Union Pacific Railroad





Ted Dunning
<tdunning@veoh.co
m> To
<hadoop-user@lucene.apache.org>
10/16/2007 12:53 cc
PM
Subject
Re: HDFS vs. CIFS
Please respond to
hadoop-user@lucen
e.apache.org









Apologies off-list. That wasn't intended to be rude.

(binary
store



. This message and any attachments contain information from Union Pacific which may be confidential and/or privileged.
If you are not the intended recipient, be aware that any disclosure, copying, distribution or use of the contents of this message is strictly prohibited by law. If you receive this message in error, please contact the sender immediately and delete the message and any attachments.

I recommend you look at MogileFS as well.

The intended usage is very different from hadoop. Hadoop is intended to
store fewer (<20 million) larger (preferably >10MB) files. It provides
strong capabilities for splitting large files across machines and hooks for
disk local computation.

Mogile is intended for large numbers of smaller files that are intended to
be stored in toto in a conventional file system on a large collection of
hosts. It provides no file splitting and no local computing hooks. There
is a framework for running scripts on large collections of files.

Neither of these exposes a normal file system, but there is a Fuse
implementation for MogileFS and an early webDAV interface for HDFS. Both of
these will be more mature over time.

Whether either is useful for you is up to you.

For that matter, you might look at KFS as well. I know nothing about it,
but the authors seem to like it. :-)

See http://blog.kosmix.com/2007/09/kosmos_filesystem_release.html

Open source is enjoyable - it's a strange world where people share what
they think - openly. It's free speech and you won't always get that from
closed source. It's often foreign for engineers who are working alone,
in well funded startups, or in established, proven, but traditional
companies.

Email doesn't convey tone very well - I think question was addressed,
and corrected you where you were possibly mistaken on a couple issues.
That's GOOD here.

I'm new to Hadoop, but not mailing lists - and this one's pretty good at
responding to technical questions, and even helping provide solutions.




TREVORSTEWART@UP.COM wrote: