Re: definitions for /proc/fs/xfs/stat

[Mark Seger <mjseger@xxxxxxxxx>]

mark - good point about the controller cache and the 1-bye/block, but remember my files in this case are only 1 block. 

dave - a long time ago I found/reported a bug in the way linux was doing their block stats.  They were basically incrementing the byte counts as block were written to cache and the reported numbers were ridiculously high.  I had wondered if something like this could be going on, but also remember the numbers being reported by xfs are much too high.  I did 10K 1K writes, which I do understand are really 1-4k page each, but that's still only 40MB.  If I add up all the 500MB/sec bytes xfs logging is reporting (even one of which is over 10 times larger), I see something on the order of of 10GB.  But again this IS with the older kernel and so may not be worth worrying about.

-mark

On Mon, Jun 17, 2013 at 8:15 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:

On Mon, Jun 17, 2013 at 10:28:19PM +0200, Stefan Ring wrote:
> > I asked around and people believe the cache is on the order of a few GB and
> > the test ran I was intentionally large enough to overshadow any cache
> > effects, running for about a minute and doing 100K 1K file creates.  The
> > disk write data was close to a sustained 475MB/sec and would have easily
> > filled the cache in the first handful of seconds which would have produced
> > enough backpressure to slow down the write rate which it never did.
>
> If a limited number of blocks gets written over and over, you won't
> ever fill the cache.

Right, and the XFS log is circular and about 1GB in size on the
filesystem under test, so should fit completely in cache. Hence
speeds faster than a physical disk are acheivable if the log stays
resident in the cache....

Cheers,

Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx