On 14/01/2014 07:49, ZHOU Yuan wrote:> Hi Loic, thanks for the education! > > I’m also trying to understand the new ‘indep’ mode. Is this new mode designed > for Ceph-EC only? It seems that all of the data in 3-copy system are > equivalent and this new algorithm should also work? >
In the best case scenario, using indep instead of firstn on replicated pools won't make a difference. However, if the crush mapper does not find the required amount of items, firstn will give ( for instance ) [1,2,4] instead of [1,2,3,4] and the replicated pool code will gracefully handle this. If using indep the result will be [1,2,CRUSH_ITEM_NONE,4] and will probably assert somewhere. Here is an example from the test suite run when you make check : https://github.com/ceph/ceph/blob/master/src/test/cli/crushtool/bad-mappings.t where 2147483647 == CRUSH_ITEM_NONE I don't know of an other reason preventing the use of indep for replicated pools. Cheers > > Sincerely, Yuan > > > On Mon, Jan 13, 2014 at 7:37 AM, Loic Dachary <l...@dachary.org > <mailto:l...@dachary.org>> wrote: > > > > On 12/01/2014 15:55, Dietmar Maurer wrote: > > From the docs: > > > > > > > > step [choose|chooseleaf] [firstn|indep] <N> <bucket-type> > > > > > > > > What exactly is the difference between ‘firstn’ and ‘indep’? > > > Hi, > > For Ceph releases up to Emperor[1], firstn is used and I'm not aware of a > use case requiring indep. As part of the effort to implement erasure coded > pools, firstn[2] and indep[3] were separated in two functions. The firstn > method is best suited for replicated pools. The indep method tries to > minimize the position changes in case an OSD becomes unavailable. For > instance, if indep finds > > [1,2,3,4] > > and after a while 3 become unavailable, it is very likely to replace it > with > > [1,2,5,4] > > It matters to erasure coded pools because > > [4,5,2,1] > > (i.e. the same OSDs but in different positions), implies more I/O. > Another difference is that in the case of a mapping failure (i.e. unable to > find the required number of OSDs), firstn will return a short list ( for > instance [1,2,3] when 4 are required ) and indep will return a list with a > placeholder at the missing position ( for instance [1,2,CRUSH_ITEM_NONE,4] ). > > Cheers > > [1] implementation in releases up to Emperor > https://github.com/ceph/ceph/blob/v0.72/src/crush/mapper.c#L295 > [2] firstn https://github.com/ceph/ceph/blob/v0.74/src/crush/mapper.c#L295 > [3] indep https://github.com/ceph/ceph/blob/v0.74/src/crush/mapper.c#L459 > > -- > Loïc Dachary, Artisan Logiciel Libre > > > _______________________________________________ > ceph-users mailing list > ceph-users@lists.ceph.com <mailto:ceph-users@lists.ceph.com> > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > -- Loïc Dachary, Artisan Logiciel Libre
signature.asc
Description: OpenPGP digital signature
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
