[Ocfs2-users] GFS

[Kurt Hackel]

> ocfs2 looks dead. the release plan is outdated and questions  
> in this list are not answered. :-( 
>  
> Jan Pilawa. 

Jan,

Your comments are dangerously misinformed.  Yes, you did post one time
three weeks ago to this list and noone responded.  Sorry about that.
Please repost your question to the list.

To understand why your post was missed all you needed to do was look at
the ocfs2-commits list archives.  There you would have seen that ocfs2
has been getting 3 to 4 commits PER DAY since the beginning of the year,
and 46 since the beginning of March.  We've been quite busy, so to say
that "ocfs2 looks dead" is just false.  In the future, please do not
respond on behalf of ocfs2 release questions.

> > Can anyone shed light on Linux's Global File System in RHEL 4 please?  
> > It looks like Oracle's OCFS.  Is there any relationship between them?  
> > If they are different things, has anyone done any performance 
> > comparison please? 

GFS, OCFS and OCFS v2 are all completely different products.  

GFS is a commercial general-purpose clustered filesystem that includes a 
stack of other cluster services.  At this time, Red Hat charges extra
for support of this product, but I do not have the specifics on cost.  

OCFS (version 1) is the original clustered filesystem put out by Oracle.
Oracle does not charge extra for support of this product if you maintain
a database support license.  See http://oss.oracle.com/projects/ocfs/
for more information on support of version 1.  It is suitable mainly for 
Oracle database files, and it performs *better* than a raw filesystem for 
such work.  It is *very* slow as a general-purpose filesystem.  I believe 
there has been a performance study done between GFS and OCFS with respect 
to database IO.  Perhaps someone can post that here?  At this time there 
is no plan to port OCFS version 1 to the 2.6 kernel (RHEL 4/SLES 9).

OCFS v2 is in current (heavy) development and is ultimately supposed to
replace the original OCFS filesystem.  It has been primarily developed
on the 2.6 kernels, and we intend to ship the product for RHEL 4, SLES 9
and RHEL 3, at least.  The goal of this product was to retain the
database IO performance of the original product while improving the
general-purpose filesystem operations to something close to a local
filesystem.  We wanted to be able to support, for example, installing an
ORACLE_HOME onto a clustered filesystem and have that be shared by
multiple nodes. This required essentially a rewrite of most of the original
codebase, including: the on-disk filesystem layout, allocation of data
and metadata, consistent journalling with jbd, proper POSIX compliance
in each filesystem syscall, and major changes to inter-node locking.

All of this work is now complete in the filesystem, and the last
remaining work is all in the small cluster stack that we implemented to
get proper locking between nodes.  The work is primarily focused now on
the lock manager itself, and userland integration with the cluster
stack.  The filesystem module has not changed significantly in months
and is relatively stable.  After the current work is complete, there
will be a substantial internal testing cycle, and then we will put out
some form of a pre-release for external testing.  I cannot give specific
dates at this point.

Once the product is deemed "stable", someone is likely to do a
performance study which compares at least GFS vs. OCFS v2 for various
operations.  This may be done by Oracle or by some outside group.
Within the development group we normally do informal comparisons which
we may share with the list, but these would not be publishable results.
Hopefully we can soon give some rough numbers for things like database
IO performance and time-to-install an Oracle home.


Thanks!
-kurt

Kurt C. Hackel
Oracle