Message-ID: <20100618152214.GP27466@think>
Date:	Fri, 18 Jun 2010 11:22:14 -0400
From:	Chris Mason <chris.mason@...cle.com>
To:	Christian Stroetmann <stroetmann@...olab.com>
Cc:	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	linux-fsdevel@...r.kernel.org, linux-btrfs@...r.kernel.org
Subject: Re: Btrfs: broken file system design (was Unbound(?) internal
 fragmentation in Btrfs)

On Fri, Jun 18, 2010 at 05:21:21PM +0200, Christian Stroetmann wrote:
> Chris Mason wrote:
> >On Fri, Jun 18, 2010 at 03:32:16PM +0200, Edward Shishkin wrote:
> >>Mat wrote:
> >>>On Thu, Jun 3, 2010 at 4:58 PM, Edward Shishkin<edward@...hat.com>  wrote:
> >>>>Hello everyone.
> >>>>
> >>>>I was asked to review/evaluate Btrfs for using in enterprise
> >>>>systems and the below are my first impressions (linux-2.6.33).
> >>>>
> >>>>The first test I have made was filling an empty 659M (/dev/sdb2)
> >>>>btrfs partition (mounted to /mnt) with 2K files:
> >>>>
> >>>># for i in $(seq 1000000); \
> >>>>do dd if=/dev/zero of=/mnt/file_$i bs=2048 count=1; done
> >>>>(terminated after getting "No space left on device" reports).
> >>>>
> >>>># ls /mnt | wc -l
> >>>>59480
> >>>>
> >>>>So, I got the "dirty" utilization 59480*2048 / (659*1024*1024) = 0.17,
> >>>>and the first obvious question is "hey, where are other 83% of my
> >>>>disk space???" I looked at the btrfs storage tree (fs_tree) and was
> >>>>shocked with the situation on the leaf level. The Appendix B shows
> >>>>5 adjacent btrfs leafs, which have the same parent.
> >>>>
> >>>>For example, look at the leaf 29425664: "items 1 free space 3892"
> >>>>(of 4096!!). Note, that this "free" space (3892) is _dead_: any
> >>>>attempts to write to the file system will result in "No space left
> >>>>on device".
> >There are two easy ways to fix this problem.  Turn off the inline
> >extents (max_inline=0) or allow splitting of the inline extents.  I
> >didn't put in the splitting simply because the complexity was high while
> >the benefits were low (in comparison with just turning off the inline
> >extents).
> But then the benefits of splitting must be high, because it solves
> this problem if inline extents are turned on.

It depends, we might also argue that for larger inline extents like this
we are better off with much larger leaves or with using max_inline=0
instead.

> >>It must be a highly unexpected and difficult question for file system
> >>developers: "how efficiently does your file system manage disk space"?
> >>
> >>In the meanwhile I confirm that Btrfs design is completely broken:
> >>records stored in the B-tree differ greatly from each other (it is
> >>unacceptable!), and the balancing algorithms have been modified in
> >>insane manner. All these factors has led to loss of *all* boundaries
> >>holding internal fragmentation and to exhaustive waste of disk space
> >>(and memory!) in spite of the property "scaling in their ability to
> >>address large storage".
> >>
> >>This is not a large storage, this is a "scalable sieve": you can not
> >>rely on finding there some given amount of water even after infinite
> >>increasing the size of the sieve (read escalating the pool of Btrfs
> >>devices).
> >>
> >>It seems that nobody have reviewed Btrfs before its inclusion to the
> >>mainline. I have only found a pair of recommendations with a common
> >>idea that Btrfs maintainer is "not a crazy man". Plus a number of
> >>papers which admire with the "Btrfs phenomena". Sigh.
> >>
> >>Well, let's decide what can we do in current situation..
> >>The first obvious point here is that we *can not* put such file system
> >>to production. Just because it doesn't provide any guarantees for our
> >>users regarding disk space utilization.
> >Are you basing all of this on inline extents?  The other extents of
> >variable size are more flexible (taking up the room in the leaf), but
> >they can also easy be split during balancing.
> If we have to split everywhere, hasn't it then some (dramatic)
> impact on the performance of the Btrfs filesystem?
> As it was said above: splitting has a high complexity.

Yes.  Both Edward and I have worked on the reiserfs sources where inline
extents were split during balancing.  It made for a few surprises in the
file read/write code.  They aren't impossible by any means, but I wanted
to avoid them.

-chris

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists