Re: How to create bcachefs?

["Marcin Mirosław" <marcin@mejor.pl>]

W dniu 26.08.2016 o 03:06, Christopher James Halse Rogers pisze:
Hi!


> On Thu, Aug 25, 2016 at 10:11 PM, Marcin Mirosław <marcin@mejor.pl> wrote:
>> W dniu 25.08.2016 o 13:09, Christopher James Halse Rogers pisze:
>>>
>>>
>>>  On Thu, Aug 25, 2016 at 7:21 PM, Marcin Mirosław <marcin@mejor.pl>
>>> wrote:
>>>>  W dniu 25.08.2016 o 02:03, Christopher James Halse Rogers pisze:
>>>>
>>>>  Hi!
>>>>
>>>>>   On Thu, Aug 25, 2016 at 7:21 AM, marcin@mejor.pl wrote:
>>>>  [...]
>>>>>>   Does it means that cache is unavailable and only tiering will be in
>>>>>>   bcachefs?
>>>>>>   And... How to mount tiered FS? When I pass one device in mount I'm
>>>>>>   getting:
>>>>>>    bcache: bch_open_as_blockdevs() register_cache_set err
>>>>>> insufficient
>>>>>>   devices
>>>>>
>>>>>   Tiering gets you all the advantages of caching, plus you can
>>>>> (with some
>>>>>   effort) have the combined filesystem size be the sum of the SSD +
>>>>> HDD
>>>>>   capacities, rather than the capacity be determined solely by the
>>>>>   capacity of the slow tier (this is not currently the case for
>>>>>  bcachefs).
>>>>
>>>>  I think that cacheing has at least such advantages over tiering:
>>>>  - allow fast read and write to files compressed with slow alghoritm
>>>>  (gzip)
>>>
>>>  I think this is getting into ā€œwhat should we call this thingā€
>>>  arguments. A naive cache is just going to promote the gzipped data to
>>>  the fast storage. On the other end, there's nothing much preventing a
>>>  sophisticated tiering system from compressing/decompressing as a
>>> part of
>>>  tier demotion/promotion.
>>
>> When (de|re)compressing would be part of demotion or promotion then I
>> agree, cache device doesn't have advantages.
> 
> You also don't get (de|re)compression for free in a caching strategy
> depending on where you put your cache. For example, a traditional bcache
> cache backing a compressed filesystem will only ever see compressed data.
> 
>>
>>>>  - can be optimized for using SSD drives
>>>
>>>  It's not clear to me how? Tiering and caching are doing the same
>>> sort of
>>>  things.
>> I don't know how bcachefs handle SSD drives. SSD are faster in
>> sequential write, I'm not sure if it could be achieved using tier
>> device. Using cache device on disk format can be different than using in
>> bcachefs and can turn random writes into sequential writes.
> 
> I believe you mean *HDDs* are faster in sequential write? This is
> exactly what tiering gets you - all writes go to the fast, SSD storage¹,
> and then data are migrated off to the slower HDD in big, sequential chunks.


SSD deals great with random reads but both SSD and HDD don't like random
writes.
I'm pasting a few links, you can read about problems with SSD
performance. Also SSD performance is going down after a dozens minutes
of intensive writes.

https://github.com/akiradeveloper/dm-writeboost
https://www.usenix.org/conference/fast12/sfs-random-write-considered-harmful-solid-state-drives
http://codecapsule.com/2014/02/12/coding-for-ssds-part-5-access-patterns-and-system-optimizations/
http://www.tomsitpro.com/articles/enterprise-ssd-testing,2-863.html


> ¹: Indeed, IIRC this is the only dataloss bug I've found in bcachefs - I
> wrote too much data too fast, the migration from tier 0 to tier 1
> couldn't keep up, and tier 0 filled up to the point where it could no
> longer write necessary journal entries, resulting in the filesystem
> being unmountable.