I am using ZFS on linux and I am confused on why the actual disk usage of files (as reported by 'du') seems so all over the place.

I created a pool called 'vault' on a hardware Dell PERC RAID (just /dev/sdb) taking all defaults except autoexpand 'on'

I then created a volume on it using

-o reserv=2040G -o quota=2040G -o recsize=4k -o acltype=posixacl

I then rsync'ed a ext4 volume over to it. For example on this volume there are two data files (Matlab *.mat files) that are 13104 and 11264 bytes in size. On the ext4 file system du on these files says 16K and 12K respectively corresponding to the 4K block size. Any <4K file will always say 4K from du.

In contrast on ZFS a du on these two files shows 25K and 21K respectively while on one 1 byte file I get 4.5K. The extra .5K on the later is not too alarming due to various metadata usage I guess. Some other <4K files I have run du on though come back exactly 4K. Most confusing is why are the du's on the *.mat files nearly twice the "real" data size?

  • I believe it has something to do with how copy-on-write filesystems like ZFS work, which can make determining actual disk usage with standard tools less reliable as with other filesystem types, due to snapshots and file revision copies and metadata. – Mio Rin Sep 7 '18 at 22:07

You can use zdb to determine information like this - just grab the inode and the dataset. For example, if your dataset is named tank/foo, you can use ls -i to determine the inode number, and then zdb -ddddd tank/foo $INODE to dump out the information.

Here's an example on my machine:

# cd /var/tmp
# mkfile 13104 file1
# ls -i file1
  4125 file1
# zdb -ddddd rpool/VARSHARE/tmp 4125
Dataset rpool/VARSHARE/tmp [ZPL], ID 4128, cr_txg 928175, 8.24G, 1223 objects, rootbp DVA[0]=<0:1f62907a00:200:STD:1> DVA[1]=<0:f4a2eda00:200:STD:1> [L0 DMU objset] fletcher4 lzjb LE unique unencrypted size=800L/200P birth=14962025L/14962025P fill=1223 contiguous 2-copy cksum=1b152e5f60:7b661ed2ecb:1445f97091785:278e0e37baf85b

    Object  lvl   iblk   dblk  dsize  lsize   %full  type
      4125    1    16K  13.0K  13.0K  13.0K  100.00  ZFS plain file
                                        168   bonus  System attributes
        dnode maxblkid: 0
        path    /file1
        uid     0
        gid     0
        atime   Fri Sep  7 18:48:00 2018
        mtime   Fri Sep  7 18:48:00 2018
        ctime   Fri Sep  7 18:48:00 2018
        crtime  Fri Sep  7 18:48:00 2018
        gen     14962023
        mode    0100600
        size    13104
        parent  4
        links   1
        pflags  0x40800000204
Indirect blocks:
                 0 L0 0:0x1f55eb0200:0x3400 0x3400L/0x3400P F=1 B=14962023/14962023 ---

                segment [000000000000000000, 0x0000000000003400) size 13.0K

This will let you know the size of the data, and the amount of metadata (labelled "indirect blocks") that your file is consuming.

In the case here, it allocated a block of 13k exactly, and is using a single indirect block of 16k. So, it's using 29k to store the 13k file. I'm guessing your numbers will be similar.

Note that the 16k "iblk" is very probably compressed, so it's a good bet it's physically taking up only 4k.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.