ZFS and large directories - caveat report

[Ivan Voras]

On 21 July 2011 20:25, Martin Matuska <mm at freebsd.org> wrote:
> Quoting:
> ... The default record size ZFS utilizes is 128K, which is good for many
> storage servers that will harbor larger files. However, when dealing with
> many files that are only a matter of tens of kilobytes, or even bytes,
> considerable slowdown will result. ZFS can easily alter the record size of
> the data to be written through the use of attributes. These attributes can
> be set at any time through the use of the "zfs set" command. To set the
> record size attribute perform "zfs set recordsize=32K pool/share". This will
> set the recordsize to 32K on share "share" within pool "pool". This type of
> functionality can even be implemented on nested shares for even more
> flexibility. ...
>
> Read more:
> http://www.articlesbase.com/information-technology-articles/improving-file-system-performance-utilizing-dynamic-record-sizes-in-zfs-4565092.html#ixzz1SlWZ7BM5

Thank you very much - now if only you took as much effort to explain
the possible connection between your quote and my post as it took you
to find the quote :)

As others explained, ZFS definitely does not use fixed block sizes,
which I can easily confirm for you by running iostat:

www:~> iostat 1
       tty             ad4              ad6              ad8             cpu
 tin  tout  KB/t tps  MB/s   KB/t tps  MB/s   KB/t tps  MB/s  us ni sy in id
   0    32  5.87 123  0.70   5.88 122  0.70   5.99 121  0.71   4  0  8  0 88
   0   233  1.54 171  0.26   1.54 166  0.25   1.53 172  0.26  10  0 17  1 72
   0    78  7.34 181  1.30   7.02 191  1.31   7.46 177  1.29   0  0  8  0 92
   0    78  5.50 234  1.25   5.59 232  1.26   5.27 247  1.27   0  0 10  1 89
^C

KB/t varies and is small.

Now I'm working under the hypothesis that the directory pseudo-file
itself is hugely fragmented and ZFS fragments it even more every time
it adds or removes from it.

Any ideas how to verify this?