Le Wed, 14 Nov 2018 17:08:59 +0100
Vojtech Pavlik <vojtech@xxxxxxxx> écrivait:
> On Wed, Nov 14, 2018 at 04:09:51PM +0100, Emmanuel Florac wrote:
>
> > am I right to assume that it's safer to create a RAID-1 (for
> > instance with md) of two SSDs before setting up bcache as a
> > writeback device?
> >
> > Any hint welcome :)
>
> That's a big question.
>
> In my opinion the safest way is to set the target for dirty data to
> zero (and reasonable writeback rate parameters).
>
> In that case, anytime the system is idle, everything gets written to
> the backing device and a loss of the caching device is not a problem.
>
Sure but caching devices may fail sometimes...
> Now why I'm not saying outright "Yes, do a RAID-1":
>
> Bcache relies on barriers for cache device consistency.
>
> Without a battery-backed RAID controller, it's hard to ensure RAID
> consistency in the event of a power loss, since barriers passed down
> the device queues are not synchronized.
>
> So in the naïve approach of just sending the same stream of requests
> to each drive, one queue can progress further than the other in the
> event of a power loss, resulting in an inconsistent content,
> resulting in subsequent data corruption when reads are interleaved.
Not a problem here: these are high end SN200 NVMe drives, with
capacitor-protected cache, caching a BBU-equipped RAID controller. Belt
and suspenders at the same time :)
>
> A battery backed memory on a RAID controller card can fix that,
> keeping all written data until confirmed by the drives. But then -
> the memory is again a single point of failure. Good thing that
> memories tend to be quite reliable.
>
Yup, never ever had any single BBU fail. And I've setup a whole lot of
them...
> Without a battery-backed RAM using a software RAID (md), Linux simply
> resorts to only sending one barrier operation at a time to the
> underlying drives plus additional housekeeping like a write intent
> bitmap.
>
> This fixes the consistency problem, however comes at a significant
> performance cost. Again, something you don't want with a cache device.
>
So you're basically telling me that using md RAID 1 could possibly be
much slower than a Broadcom 94xx RAID controller for NVMe drives?
Interesting.
> So, make your own decision based on your usecase. :)
Use case is : as fast as possible, but doesn't endanger the 600 TB of
data, because duh, 600 TB is quite a lot :)
--
------------------------------------------------------------------------
Emmanuel Florac | Direction technique
| Intellique
| <eflorac@xxxxxxxxxxxxxx>
| +33 1 78 94 84 02
------------------------------------------------------------------------
Attachment:
pgpIfZVDdotz4.pgp
Description: OpenPGP digital signature
