Thank you Ilya for the detailed explanation. Regards,
Ridwan Noel On Mon, Oct 24, 2016 at 6:30 AM, Ilya Dryomov <idryo...@gmail.com> wrote: > On Fri, Oct 21, 2016 at 10:35 PM, Ridwan Rashid Noel > <ridwan...@gmail.com> wrote: > > Thank you for your reply Greg. Is there any detailed resource that > describe > > about how the primary affinity changing works? All I got from searching > was > > one paragraph from the documentation. > > No, probably nothing detailed. There isn't much to it though. > > Normally, the first up OSD in the raw OSD set returned by CRUSH is the > primary. primary-affinity then sets the probability with which the > mapping code will choose some other OSD from the raw OSD set to serve > as primary. (Your quote is technically incorrect, as this adjustment > happens after CRUSH is run, not within CRUSH.) > > You can think of it as a step in the mapping process, which will > potentially reorder OSDs in the OSD set. Say your CRUSH output is > [osd0, osd1, osd2] and all of them are up. If osd0's primary-affinity > is 1, osd0 will be the primary with p=1 (i.e. always). If osd0's > primary-affinity is 0.75, either osd1 or osd2 will take its place with > p=0.25 and the resulting OSD set will be one of: > > [osd0, osd1, osd2], p=0.75 > [osd1, osd0, osd2] or [osd2, osd0, osd1], p=0.25 > > Once you've adjusted the primary-affinity, this choice is fixed until > you either re-adjust the primary-affinity or the mapping is affected in > some other way (e.g. OSD goes down, etc). > > Another way to look at it is as follows. If, according to CRUSH + > current osdmap, osd0 gets to be the primary for 100 PGs, 25 of those > PGs will be rejected and get another primary, leaving osd0 with 75 PGs. > > Why would you want to do this? To quickly redistribute the read > workload away from overloaded OSDs (writes go to all OSDs in the PG, > while reads are served by the primary). > > Thanks, > > Ilya >
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
