Andi,
Also, Van mentioned the issue with double-caching. Since Oracle uses
it's own cache, the current buffer cache implementation in Linux for
O_SYNC obsolutley sucks for Oracle, since we will always be double
buffering data. I mentioned to Alan having a directFS layer in Linux,
where Oracle could talk to files in a way that would completely by-pass
the Linux buffer cache. Given the current state of Linux and the file
systems, this may be very difficult to support.
:-)
Jeff
"Jeff V. Merkey" wrote:
>
> Andi,
>
> Van and I worked together for several years on Oracle for NetWare (I
> guess this part was obvious). He is getting together some info for us
> on what would be the most ideal. A block oriented version of fsync() is
> probably the way to go the get Oracle the highest numbers possible on
> Linux rather then force a flush of every dirty block for a file each
> time an Oracle data table is written to.
>
> We should probably add (int block, int blocks) to fsync() to give
> callers the ability to selectively push blocks out of cache. IMHO.
>
> Jeff
>
> Andi Kleen wrote:
> >
> > "Jeff V. Merkey" <jmerkey@timpanogas.com> writes:
> > >
> > > So what's wrong with calling fsync() after write() then for Oracle --
> > > this should work -- yes?
> >
> > I would expect O_SYNC to be a bit faster than fsync: fsync has to figure
> > out what to flush, in O_SYNC it is implicit and already known.
> >
> > At least in ext2 fsync is relatively costly CPU wise [there were real
> > error reports about that in the past], while O_SYNC is much cheaper.
> >
> > This is now especially true because O_SYNC in 2.3 does not actually
> > flush. This of course needs to be changed.
> >
> > -Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon May 15 2000 - 21:00:27 EST