Hi, comments inline. > On 05 Aug 2015, at 05:45, Jevon Qiao <qiaojianfeng@xxxxxxxxxxxxxxx> wrote: > > Hi Jan, > > Thank you for the detailed suggestion. Please see my reply in-line. > On 5/8/15 01:23, Jan Schermer wrote: >> I think I wrote about my experience with this about 3 months ago, including what techniques I used to minimize impact on production. >> >> Basicaly we had to >> 1) increase pg_num in small increments only, bcreating the placement groups themselves caused slowed requests on OSDs >> 2) increse pgp_num in small increments and then go higher > So you totally completed the step 1 before jumping into step 2. Have you ever tried mixing them together? Increase pg_number, increase pgp_number, increase pg_number… Actually we first increased both to 8192 and then decided to go higher, but that doesn’t matter. The only reason for this was that the first step took could run unattended at night without disturbing the workload.* The second step had to be attended. * in other words, we didn’t see “slow requests” because of our threshold settings, but while PGs were creating the cluster paused IO for non-trivial amounts of time. I suggest you do this in as small steps as possible, depending on your SLAs. >> We went from 4096 placement groups up to 16384 >> >> pg_num (the number of on-disk created placement groups) was increased like this: >> # for i in `seq 4096 64 16384` ; do ceph osd pool set $pool pg_num $i ; sleep 60 ; done >> this ran overnight (and was upped to 128 step during the night) >> >> Increasing pgp_num was trickier in our case, first because it was heavy production and we wanted to minimize the visible impact and second because of wildly differing free space on the OSDs. >> We did it again in steps and waited for the cluster to settle before continuing. >> Each step upped pgp_num by about 2% and as we got higher (>8192) we increased this to much more - the last step was 15360->16384 with the same impact the initial 4096->4160 had. > The strategy you adopted looks great. I'll do some experiments on a test cluster to evaluate the real impact in each step >> The end result is much better but still nowhere near optimal - bigger impact would be upgrading to a newer Ceph release and setting the new tunables because we’re running Dumpling. >> >> Be aware that PGs cost some space (rough estimate is 5GB per OSD in our case), and also quite a bit of memory - each OSD has 1.7-2.0GB RSS right now while it only had about 1GB before. That’s a lot of memory and space with higher OSD counts... > This is a good point. So along with the increment of PGs, we also need to take the current status of the cluster(the available disk space and memory for each OSD) into account and evaluate whether it is needed to add more resources. Depends on how much free space you have. We had some OSDs at close to 85% capacity before we started (and other OSD’s at only 30%). When increasing the number of PGs the data shuffled greatly - but this depends on what CRUSH rules you have (and what version you are running). Newer versions with newer tunables will make this a lot easier I guess. >> And while I haven’t calculated the number of _objects_ per PG, but we have differing numbers of _placement_groups_ per OSD (one OSD hosts 500, another hosts 1300) and this seems to be the cause of poor data balancing. > In our environment, we also encountered the imbalance mapping between PGs and OSD. What kind of bucket algorithm was used in your environment? Any idea on how to minimize it? We are using straw because of dumpling. Straw2 should make everything better :-) Jan > > Thanks, > Jevon >> Jan >> >> >>> On 04 Aug 2015, at 18:52, Marek Dohojda <mdohojda@xxxxxxxxxxxxxxxxxxx> wrote: >>> >>> I have done this not that long ago. My original PG estimates were wrong and I had to increase them. >>> >>> After increasing the PG numbers the Ceph rebalanced, and that took a while. To be honest in my case the slowdown wasn’t really visible, but it took a while. >>> >>> My strong suggestion to you would be to do it in a long IO time, and be prepared that this willl take quite a long time to accomplish. Do it slowly and do not increase multiple pools at once. >>> >>> It isn’t recommended practice but doable. >>> >>> >>> >>>> On Aug 4, 2015, at 10:46 AM, Samuel Just <sjust@xxxxxxxxxx> wrote: >>>> >>>> It will cause a large amount of data movement. Each new pg after the >>>> split will relocate. It might be ok if you do it slowly. Experiment >>>> on a test cluster. >>>> -Sam >>>> >>>> On Mon, Aug 3, 2015 at 12:57 AM, 乔建峰 <scaleqiao@xxxxxxxxx> wrote: >>>>> Hi Cephers, >>>>> >>>>> This is a greeting from Jevon. Currently, I'm experiencing an issue which >>>>> suffers me a lot, so I'm writing to ask for your comments/help/suggestions. >>>>> More details are provided bellow. >>>>> >>>>> Issue: >>>>> I set up a cluster having 24 OSDs and created one pool with 1024 placement >>>>> groups on it for a small startup company. The number 1024 was calculated per >>>>> the equation 'OSDs * 100'/pool size. The cluster have been running quite >>>>> well for a long time. But recently, our monitoring system always complains >>>>> that some disks' usage exceed 85%. I log into the system and find out that >>>>> some disks' usage are really very high, but some are not(less than 60%). >>>>> Each time when the issue happens, I have to manually re-balance the >>>>> distribution. This is a short-term solution, I'm not willing to do it all >>>>> the time. >>>>> >>>>> Two long-term solutions come in my mind, >>>>> 1) Ask the customers to expand their clusters by adding more OSDs. But I >>>>> think they will ask me to explain the reason of the imbalance data >>>>> distribution. We've already done some analysis on the environment, we >>>>> learned that the most imbalance part in the CRUSH is the mapping between >>>>> object and pg. The biggest pg has 613 objects, while the smallest pg only >>>>> has 226 objects. >>>>> >>>>> 2) Increase the number of placement groups. It can be of great help for >>>>> statistically uniform data distribution, but it can also incur significant >>>>> data movement as PGs are effective being split. I just cannot do it in our >>>>> customers' environment before we 100% understand the consequence. So anyone >>>>> did this under a production environment? How much does this operation affect >>>>> the performance of Clients? >>>>> >>>>> Any comments/help/suggestions will be highly appreciated. >>>>> >>>>> -- >>>>> Best Regards >>>>> Jevon >>>>> >>>>> _______________________________________________ >>>>> ceph-users mailing list >>>>> ceph-users@xxxxxxxxxxxxxx >>>>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >>>>> >>>> _______________________________________________ >>>> ceph-users mailing list >>>> ceph-users@xxxxxxxxxxxxxx >>>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >>> _______________________________________________ >>> ceph-users mailing list >>> ceph-users@xxxxxxxxxxxxxx >>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >> -- >> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in >> the body of a message to majordomo@xxxxxxxxxxxxxxx >> More majordomo info at http://vger.kernel.org/majordomo-info.html > _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |