On 2015-12-02 08:53, Tomasz Chmielewski wrote:
> On 2015-12-02 22:03, Austin S Hemmelgarn wrote:
>
>>>  From these numbers (124 GB used where data size is 153 GB), it appears
>>> that we save around 20% with zlib compression enabled.
>>> Is 20% reasonable saving for zlib? Typically text compresses much better
>>> with that algorithm, although I understand that we have several
>>> limitations when applying that on a filesystem level.
>>
>> This is actually an excellent question.  A couple of things to note
>> before I share what I've seen:
>> 1. Text compresses better with any compression algorithm.  It is by
>> nature highly patterned and moderately redundant data, which is what
>> benefits the most from compression.
>
> It looks that compress=zlib does not compress very well. Following
> Duncan's suggestion, I've changed it to compress-force=zlib, and
> re-copied the data to make sure the file are compressed.
For future reference, if you run 'btrfs filesystem defrag -r -czlib' on 
the top level directory, you can achieve the same effect without having 
to deal with the copy overhead.  This has a side effect of breaking 
reflinks, but copying the files off and back onto the filesystem does so 
also, and even then, I doubt that you're using reflinks.  There probably 
wouldn't be much difference in the time it takes, but at least you 
wouldn't be hitting another disk in the process.
>
> Compression ratio is much much better now (on a slightly changed data set):
>
> # df -h
> /dev/xvdb       200G   24G  176G  12% /var/log/remote
>
>
> # du -sh /var/log/remote/
> 138G    /var/log/remote/
>
>
> So, 138 GB files use just 24 GB on disk - nice!
>
> However, I would still expect that compress=zlib has almost the same
> effect as compress-force=zlib, for 100% text files/logs.
>
That's better than 80% space savings (it works out to about 83.6%), so I 
doubt that you'd manage to get anything better than that even with only 
plain text files.  It's interesting that there's such a big discrepancy 
though, that indicates that BTRFS really needs some work WRT deciding 
what to compress.