Description of problem and Proposal of improvement:

In the kernel implementation of RHEL4 or RHEL5, 
Even if fdatasync(2) is used, it always becomes considerable fsync(2). 

fdatasync(2) (that contains only the overwrite) becomes high-speed
by not synchronizing the inode in only case of overwrite which the
meta-data doesn't change.

Then, I propose the following patches. The patch can be easily applied
even by RHEL4 and RHEL5. It has already been taken into Vanilla-kernel.
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=3d61f75eefedf75914ab4453c67aaa2ee64bcf93;hp=8b91582500ae750db22bd515379616e5e9ad06ab
> > --- a/fs/ext3/fsync.c
> > +++ b/fs/ext3/fsync.c
> > @@ -72,6 +72,9 @@ int ext3_sync_file(struct file * file, struct dentry
*dentry, int datasync)
> >                 goto out;
> >         }
> >  
> > +       if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
> > +               goto out;
> > +
> >         /*
> >          * The VFS has written the file data.  If the inode is unaltered
> >          * then we need not start a commit.

Additional info:

For PostgreSQL, the patch has a dramatic effect on to the performance gain.
Because fdatasync(2) is executed the following steps:
  1. Enhance the transaction log file (= WAL) by 0 padding.
  2. Write the contents of transaction.
  3. Synchronize by executing fdatasync(2) per a transaction.


Please take it by the following update! (RHEL5.2 or 5.3, RHEL4.7 or 4.8)