On Tue, 2009-03-17 at 15:57 -0500, Steven Pratt wrote:
> Chris Mason wrote:
> > On Sun, 2009-03-15 at 09:38 -0500, Steven Pratt wrote:
> >   
> >>> Thanks for running this, but the main performance fixes for your test
> >>> are still in testing locally.  One thing that makes a huge difference on
> >>> the random write run is to mount -o ssd.
> >>>
> >>>   
> >>>       
> >> Tried a run with -o ssd on the raid system. It made some minor 
> >> improvements in random write performance. Helps more on odirect, but 
> >> mainly at the 16thread count. Single and 128 threads it doesn't make 
> >> much difference.
> >>
> >> Results syncing now to history boxacle
> >> http://btrfs.boxacle.net/repository/raid/history/History.html
> >>     
> >
> > Well, still completely different from my test rig ;)  For the random
> > write run, yours runs at 580 trans/sec for btrfs and mine is going along
> > at 8000 trans/sec.
> >   
> That is odd.  However, I think I have found 1 factor.  In rerunning with 
> blktrace and sysrq an interesting thing happened.  The results got a lot 
> faster.  What I did was just run the 128 thread odirect random write 
> test.  Instead of 2.8MB/sec, I got 17MB/sec. Still far below the 100+ of 
> ext4 and JFS, but one heck of a difference.  Here is what I think is 
> going on.  We make use of a flag in FFSB to reuse the existing fileset 
> if the fileset meets the setup criteria exactly.  For the test I am 
> running that is 1024 100MB files.  Since all of the random write test 
> are doing overwrites within the file, the file sizes do not change and 
> therefore the fileset is valid for reuse.  

Oh! In that case you're stuck waiting to cache the extents already used
in a block group.  At least I hope that's what sysrq-w will show us.
The first mods to a block group after a mount are slow while we read in
the free extents.

-chris


--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to