Luminous addresses it with a mgr plugin that actively changes the weights of OSDs to balance the distribution. In addition to having PGs distributed well for your OSDs to have an equal amount of data on them is also which OSDs are Primary. If you're running into a lot of latency on specific OSDs during high read volumes, then it might just be that there are too many PGs that the OSD is primary for affecting it. All reads happen on the primary OSD even though you have multiple secondary OSDs for the PG. Just a thought, might not be what's happening.
On Thu, Mar 8, 2018 at 11:26 PM shadow_lin <shadow_...@163.com> wrote: > Thanks for your advice. > I will try to reweight osds of my cluster. > > Why ceph is so sensitive to unblanced pg distribution during high load? > ceph osd df result is: https://pastebin.com/ur4Q9jsA. ceph osd perf > result is: https://pastebin.com/87DitPhV > > There is no osd with very high pg count compare to others. When the wirte > test load is low everything seems fine, but during high write load test, > some of the osds with higher pg can have 3-10 time of fs_apply_latency > compare to others. > > My guess is the high loaded osds kinda slowed the whole cluster（because I > have only one pool with all osds）to the level of how fast they can handle. > So other osd has lower load and have a good latency. > > Is this expected during high load(Indicate the load is too hight for > current cluster to hanlde)? > > How does luminous solve the unevenly pg distribution problem?I read about > there is a pg-upmap exception table in the osdmap in luminous 12.2.x. It is > said to use this it is possible to achive perfect pg distribution among > osds. > > 2018-03-09 > ------------------------------ > shadow_lin > ------------------------------ > > *发件人：*David Turner <drakonst...@gmail.com> > *发送时间：*2018-03-09 06:45 > *主题：*Re: [ceph-users] Uneven pg distribution cause high fs_apply_latency > on osds with more pgs > *收件人：*"shadow_lin"<shadow_...@163.com> > *抄送：*"ceph-users"<ceph-users@lists.ceph.com> > > > PGs being unevenly distributed is a common occurrence in Ceph. Luminous > started making some steps towards correcting this, but you're in Jewel. > There are a lot of threads in the ML archives about fixing PG > distribution. Generally every method comes down to increasing the weight > on OSDs with too few PGs and decreasing the weight on the OSDs with too > many PGs. There are a lot of schools of thought on the best way to > implement this in your environment which has everything to do with your > client IO patterns and workloads. Looking into `ceph osd reweight-by-pg` > might be a good place for you to start as you are only looking at 1 pool in > your cluster. If you have more pools, you generally need `ceph osd > reweight-by-utilization`. > > On Wed, Mar 7, 2018 at 8:19 AM shadow_lin <shadow_...@163.com> wrote: > >> Hi list, >> Ceph version is jewel 10.2.10 and all osd are using filestore. >> The Cluster has 96 osds and 1 pool with size=2 replication with 4096 >> pg(base on pg calculate method from ceph doc for 100pg/per osd). >> The osd with the most pg count has 104 PGs and there are 6 osds have >> above 100 PGs >> Most of the osd have around 7x-9x PGs >> The osd with the least pg count has 58 PGs >> >> During the write test some of the osds have very high fs_apply_latency >> like 1000ms-4000ms while the normal ones are like 100-600ms. The osds with >> high latency are always the ones with more pg on it. >> >> iostat on the high latency osd shows the hdds are having high %util at >> about 95%-96% while the normal ones are having %util at 40%-60% >> >> I think the reason to cause this is because the osds have more pgs need >> to handle more write request to it.Is this right? >> But even though the pg distribution is not even but the variation is not >> that much.How could the performance be so sensitive to it? >> >> Is there anything I can do to improve the performance and reduce the >> latency? >> >> How can I make the pg distribution to be more even? >> >> Thanks >> >> >> 2018-03-07 >> ------------------------------ >> shadowlin >> >> _______________________________________________ >> ceph-users mailing list >> ceph-users@lists.ceph.com >> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >> >
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
