RAID-Z wasted space - asize roundups to nparity +1

[Adam Nowacki]

On 2013-01-28 22:55, Matthew Ahrens wrote:
> This is so that we won't end up with small, unallocatable segments.
>   E.g. if you are using RAIDZ2, the smallest usable segment would be 3
> sectors (1 sector data + 2 sectors parity).  If we left a 1 or 2 sector
> free segment, it would be unusable and you'd be able to get into strange
> accounting situations where you have free space but can't write because
> you're "out of space".

Sounds reasonable.

> The amount of waste due to this can be minimized by using larger
> blocksizes (e.g. the default recordsize of 128k and files larger than
> 128k), and by using smaller sector sizes (e.g. 512b sector disks rather
> than 4k sector disks).  In your case these techniques would limit the
> waste to 0.6%.

This brings another issue - recordsize capped at 128KiB. We are using 
the pool for off-line storage of large files (from 50MB to 20GB). Files 
are stored and read sequentially as a whole. With 12 disks in RAID-Z2, 
4KiB sectors, 128KiB record size and the padding above 9.4% of disk 
space goes completely unused - one whole disk.

Increasing recordsize cap seems trivial enough. On-disk structures and 
kernel code support it already - a single of code had to be changed 
(#define SPA_MAXBLOCKSHIFT - from 17 to 20) to support 1MiB recordsizes. 
This of course breaks compatibility with any other system without this 
modification. With Suns cooperation this could be handled in safe and 
compatible manner via pool version upgrade. Recordsize of 128KiB would 
remain the default but anyone could increase it with zfs set.

Pool appears to work just fine with 15TB copied so far from another 
pool. Wasted disk space drops down to 0.7%. Sequential read speed 
increased from ~400MB/s to ~600MB/s. Writes stay about the same at ~300MB/s.

So far however I was not able to boot from that pool. gptzfsboot 
required a heap size increase and appears to work. zfsloader crashes and 
I've become lost in the code.

I've also identified another problem with ZFS wasting disk space. When 
compression is off allocations are always a multiple of record size. 
With the default recordsize of 128KiB a 129KiB file would use 256KiB of 
disk space (+ parity and other inefficiencies mentioned above). This may 
be there to help with fragmentation but then it would be good to have a 
setting to turn it off - even if by means of a no-op compression that 
would count zeroes backwards and return short psize.

>
> --matt
>
> On Sun, Jan 27, 2013 at 5:01 AM, Adam Nowacki <nowakpl at platinum.linux.pl
> <mailto:nowakpl at platinum.linux.pl>> wrote:
>
>     I've just found something very weird in the ZFS code.
>
>     sys/cddl/contrib/opensolaris/__uts/common/fs/zfs/vdev_raidz.__c:504
>     in HEAD
>
>     Can someone explain the reason behind this line of code? What it
>     does is align on-disk record size to a multiple of number of parity
>     disks + 1 ... this really doesn't make any sense. So far as I can
>     tell those extra sectors are just padding - completely unused.
>
>     For the array I'm using this results in 4.8% of wasted disk space -
>     1.7TB. It's a 12x 3TB disk RAID-Z2.
>     _________________________________________________
>     freebsd-fs at freebsd.org <mailto:freebsd-fs at freebsd.org> mailing list
>     http://lists.freebsd.org/__mailman/listinfo/freebsd-fs
>     <http://lists.freebsd.org/mailman/listinfo/freebsd-fs>
>     To unsubscribe, send any mail to
>     "freebsd-fs-unsubscribe at __freebsd.org
>     <mailto:freebsd-fs-unsubscribe at freebsd.org>"
>
>