Hi Frank, Just a guess, but I wonder if for small values rounding/precision start to impact the placement like you observed.
Do you see the same issue if you reweight to 2x the original? -- Dan On Tue, Nov 15, 2022 at 10:09 AM Frank Schilder <fr...@dtu.dk> wrote: > > Hi all, > > I re-weighted all OSDs in a pool down from 1.0 to the same value 0.052 (see > reason below). After this, all hell broke loose. OSDs were marked down, slow > OPS all over the place and the MDSes started complaining about slow > ops/requests. Basically all PGs were remapped. After setting all re-weights > back to 1.0 the situation went back to normal. > > Expected behaviour: No (!!!) PGs are remapped and everything continues to > work. Why did things go down? > > More details: We have 24 OSDs with weight=1.74699 in a pool. I wanted to add > OSDs with weight=0.09099 in such a way that the small OSDs receive > approximately the same number of PGs as the large ones. Setting a re-weight > factor of 0.052 for the large ones should achieve just that: > 1.74699*0.05=0.09084. So, procedure was: > > - ceph osd crush reweight osd.N 0.052 for all OSDs in that pool > - add the small disks and re-balance > > I would expect that the crush mapping is invariant under a uniform change of > weight. That is, if I apply the same relative weight-change to all OSDs > (new_weight=old_weight*common_factor) in a pool, the mappings should be > preserved. However, this is not what I observed. How is it possible that PG > mappings change if the relative weight of all OSDs to each other stays the > same (the probabilities of picking an OSD are unchanged over all OSDs)? > > Thanks for any hints. > > Best regards, > ================= > Frank Schilder > AIT Risø Campus > Bygning 109, rum S14 > _______________________________________________ > ceph-users mailing list -- ceph-users@ceph.io > To unsubscribe send an email to ceph-users-le...@ceph.io _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io