On 2015-06-15 10:44, Tovo Rabemanantsoa wrote:
> On 06/15/2015 03:29 PM, David Sterba wrote:
>> On Mon, Jun 15, 2015 at 02:47:24PM +0200, Tovo Rabemanantsoa wrote:
>>> Hi all,
>>> By browsing this list's archive, I've found a thread initiated by
>>> Charles Cazabon entitled: "Oddly slow read performance with near-full
>>> largish FS."
>>> Actually, I'm living the same experience but with a not so large FS
>>> (256GB on a SSD). Indeed, when I have less than 1GB of free space, the
>>> applications (thunderbird, thunar ...) on the machine become awfully
>>> slow but remain normal if I make some cleaning.
>>> Is it due to the FS or because it's an SSD hard disk ?
>>
>> 1G of 256G is less than a percent. At this level of usage you can expect
>> slowdown on any filesystem.
>>
>> This could be caused by free space fragmentation and even on a SSD, this
>> needs extra time to process.  Higher number of fragments needs more
>> structures to represent them and cost more CPU time, though this still
>> might not be the worst impact.
>>
>> AFAIK btrfs space handling logic needs to do more flushes of unwritten
>> data when the accounted free space goes below some threshold (because
>> COW needs to write the data twice before it switches to the new "root"
>> pointer and can free the previous version).
>
> Thanks for you reply,
> If I really understand, it's always a good idea to keep more than 1% of
> free space. Right ?
For almost any non-COW filesystem (ext4, XFS, JFS, etc.), 1% or 100MB 
(whichever is larger) is generally a good buffer.  On BTRFS, I would say 
at least 5% or 1.5G (again, whichever is larger; and if performance is a 
concern, go for at least 10-20%), as BTRFS is known to have some rather 
poor behavior when running very close to full.