Re: ext4 features (checksums)

[Avi Kivity <avi@argo.co.il>]

Neil Brown wrote:
> On Tuesday July 4, avi@argo.co.il wrote:
> > Neil Brown wrote:
> > >
> > > To my mind, the only thing you should put between the filesystem and
> > > the raw devices is RAID (real-raid - not raid0 or linear).
> > >
> > I believe that implementing RAID in the filesystem has many benefits 
> too:
> >  - multiple RAID levels: store metadata in triple-mirror RAID 1, random
> > write intensive data in RAID 1, bulk data in RAID 5/6
> >  - improved write throughput - since stripes can be variable size, any
> > large enough write fills a whole stripe
>
> Maybe....
>
> Now imagine what would be required to rebuild a whole drive onto a
> spare after a drive failure.
>
> I'm sure it is possible, and I believe ZFS does something like that.
> I find it hard to imagine getting reasonable speed if there is much
> complexity.  And the longer it takes, the longer your data is exposed
> to multiple-failures.

A company called Isilon does this on a cluster.  They claim (IIRC) a one 
hour rebuild time for a failure.  AFAIK they rebuild into cluster free 
space, so they are not bound by the spare's bandwidth; they can utilize 
all cluster resources for a rebuild.

(You don't need spare disks, just spare free space; so you don't have 
idle disk heads)

In terms of complexity, I imagine one needs a reverse mapping (extent -> 
(inode, offset)); given that, one can very easily rebuild failed disks, 
and more features are easy to implement, like evacuation of a drive, or 
rebalancing data across all drives when new disks are added.

The same ideas can be applied to a non-clustered filesystem, of course.

--
error compiling committee.c: too many arguments to function

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/