Re: [RFC v3] ext4: Combine barrier requests coming from fsync

["Darrick J. Wong" <djwong@us.ibm.com>]

On Mon, Aug 16, 2010 at 09:14:33AM -0700, Darrick J. Wong wrote:
> On Mon, Aug 09, 2010 at 05:07:23PM -0400, Christoph Hellwig wrote:
> > Can you try with the new barrier implementation in the
> > 
> > 	[PATCH, RFC] relaxed barriers
> > 
> > by making cache flushes just that and not complicated drain barrier
> > it should speed this case up a lot.
> 
> Indeed it does!  The barrier count increases to about 21000, but I also see
> much higher throughput, about 830 transactions per second (versus 12000 and 760
> respectively before Tejun's patch).

Oddly, I ran the entire suite of tests against a larger set of machines, and
with Tejun's RFC patchset I didn't see nearly as much of an improvement.  I
have been trying to put together a new tree based on "replace barrier with
sequenced flush" and Christoph's "explicitly flush/FUA" patch sets, though I've
gotten lost in the weeds. :(

I also experienced some sort of crash with Tejun's relaxed barrier patch on one
of my systems.  I was hitting the BUG_ON in drivers/scsi/scsi_lib.c, line 1115.

Rather than hold on to (aging) test results any further, I'll be posting a new
fsync coordination patch shortly that includes Andreas' suggestion to use the
average barrier time instead of a static 500us, and a spreadsheet that shows
what happens with various patches, and on a wider range of hardware.

--D