Carsten Aulbert wrote:
Hi all,
I was just reading
http://blogs.sun.com/dap/entry/zfs_compression
and would like to know what the experience of people is about enabling
compression in ZFS.
In principle I don't think it's a bad thing, especially not when the
CPUs are fast enough to improve the performance as the hard drives might
be too slow. However, I'm missing two aspects:
o what happens when a user opens the file and does a lot of seeking
inside the file? For example our scientists use a data format where
quite compressible data is contained in stretches and the file header
contains a dictionary where each stretch of data starts. If these files
are compressed on disk, what will happen with ZFS? Will it just make
educated guesses, or does it have to read all of the typically 30-150 MB
of the file and then does the seeking from buffer caches?
Files are not compressed in ZFS. Blocks are compressed.
If the compression of the blocks cannot gain more than 12.5% space savings,
then the block will not be compressed. If your file contains
compressable parts
and uncompressable parts, then (depending on the size/blocks) it may be
partially compressed.
o Another problem I see (but probably isn't): A user is accessing a file
via a NFS-exported ZFS, appending a line of text, closing the file (and
hopefully also flushing everything correctly. However, then the user
opens it again appends another line of text, ... Imagine this happening
a few times per second. How will ZFS react to this pattern? Will it only
opens the final record of the file, uncompress it, adds data,
recompresses it, flushes it to disk and reports that back to the user's
processes? Is there a potential problem here?
The file will be cached in RAM. When the file is closed and synced, the data
will be written to the ZIL and ultimately to the data set. I don't
think there
is a fundamental problem here... you should notice the NFS sync behaviour
whether the backing store is ZFS or some other file system. Using a slog
or nonvolatile write cache will help performance for such workloads.
Cheers (and sorry if these questions are stupid ones)
They are good questions :-)
-- richard
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
