Sorry to steal it, so if I have 500GB and 700GB mixed wal+rocksdb on nvme the number should be the level base 5000000000 and 7000000000? Or needs to be power of 2? Istvan Szabo Senior Infrastructure Engineer --------------------------------------------------- Agoda Services Co., Ltd. e: istvan.szabo@xxxxxxxxx<mailto:istvan.szabo@xxxxxxxxx> --------------------------------------------------- On 2021. Sep 21., at 9:19, Christian Wuerdig <christian.wuerdig@xxxxxxxxx> wrote: Email received from the internet. If in doubt, don't click any link nor open any attachment ! ________________________________ It's been discussed a few times on the list but RocksDB levels essentially grow by a factor of 10 (max_bytes_for_level_multiplier) by default and you need (level-1)*10 space for the next level on your drive to avoid spill over So the sequence (by default) is 256MB -> 2.56GB -> 25.6GB -> 256GB GB and since 50GB < 286 (sum of all levels) you get spill-over going from L2 to L3. See also https://docs.ceph.com/en/latest/rados/configuration/bluestore-config-ref/#sizing Interestingly your level base seems to be 512MB instead of the default 256MB - did you change that? In your case the sequence I would have expected is 0.5 -> 5 -> 50 - and you should have already seen spillover at 5GB (since you only have 50GB partitions but you need 55.5GB at least) Not sure what's up with that. I think you need to re-create OSDs after changing these RocksDB params Overall since Pacific this no longer holds entirely true since RocksDB sharding was added ( https://docs.ceph.com/en/latest/rados/configuration/bluestore-config-ref/#bluestore-rocksdb-sharding) - it was broken in 16.2.4 but looks like it's fixed in 16.2.6 1. Upgrade to Pacific 2. Get rid of the NVME raid 3. Make 160GB DB partitions 4. Activate RocksDB sharding 5. Don't worry about RocksDB params If you don't feel like upgrading to Pacific any time soon but want to make more efficient use of the NVME and don't mind going out on a limp I'd still do 2+3 plus study https://github-wiki-see.page/m/facebook/rocksdb/wiki/Leveled-Compaction carefully and make adjustments based on that. With 160GB partitions a multiplier of 7 might work well with a base size of 350MB 0.35 -> 2.45 -> 17.15 -> 120.05 (total 140GB out of 160GB) You could also try to switch to a 9x multiplier and re-create one of the OSDs to see how it pans out prior to dissolving the raid1 setup (given your settings that should result in 0.5 -> 4.5 -> 40.5 GB usage) On Tue, 21 Sept 2021 at 13:19, mhnx <morphinwithyou@xxxxxxxxx> wrote: Hello everyone! I want to understand the concept and tune my rocksDB options on nautilus 14.2.16. osd.178 spilled over 102 GiB metadata from 'db' device (24 GiB used of 50 GiB) to slow device osd.180 spilled over 91 GiB metadata from 'db' device (33 GiB used of 50 GiB) to slow device The problem is, I have the spill over warnings like the rest of the community. I tuned RocksDB Options with the settings below but the problem still exists and I wonder if I did anything wrong. I still have the Spill Overs and also some times index SSD's are getting down due to compaction problems and can not start them until I do offline compaction. Let me tell you about my hardware right? Every server in my system has: HDD - 19 x TOSHIBA MG08SCA16TEY 16.0TB for EC pool. SSD - 3 x SAMSUNG MZILS960HEHP/007 GXL0 960GB NVME - 2 x PM1725b 1.6TB I'm using Raid 1 Nvme for Bluestore DB. I dont have WAL. 19*50GB = 950GB total usage on NVME. (I was thinking use the rest but regret it now) So! Finally let's check my RocksDB Options: [osd] bluefs_buffered_io = true bluestore_rocksdb_options = compression=kNoCompression,max_write_buffer_number=32,min_write_buffer_number_to_merge=2,recycle_log_file_num=32,compaction_style=kCompactionStyleLevel,write_buffer_size=67108864,target_file_size_base=67108864,max_background_compactions=31,level0_file_num_compaction_trigger=8,level0_slowdown_writes_trigger=32,level0_stop_writes_trigger=64,flusher_threads=8,compaction_readahead_size=2MB,compaction_threads=16, *max_bytes_for_level_base=536870912*, *max_bytes_for_level_multiplier=10* *"ceph osd df tree" *to see ssd and hdd usage, omap and meta. ID CLASS WEIGHT REWEIGHT SIZE RAW USE DATA OMAP META AVAIL %USE VAR PGS STATUS TYPE NAME -28 280.04810 - 280 TiB 169 TiB 166 TiB 688 GiB 2.4 TiB 111 TiB 60.40 1.00 - host MHNX1 178 hdd 14.60149 1.00000 15 TiB 8.6 TiB 8.5 TiB 44 KiB 126 GiB 6.0 TiB 59.21 0.98 174 up osd.178 179 ssd 0.87329 1.00000 894 GiB 415 GiB 89 GiB 321 GiB 5.4 GiB 479 GiB 46.46 0.77 104 up osd.179 I know the size of NVME is not suitable for 16TB HDD's. I should have more but the expense is cutting us pieces. Because of that I think I'll see the spill overs no matter what I do. But maybe I will make it better with your help! *My questions are:* 1- What is the meaning of (33 GiB used of 50 GiB) 2- Why it's not 50GiB / 50GiB ? 3- Do I have 17GiB unused area on the DB partition? 4- Is there anything wrong with my Rocksdb options? 5- How can I be sure and find the good Rocksdb Options for Ceph? 6- How can I measure the change and test it? 7- Do I need different RocksDB options for HDD's and SSD's ? 8- If I stop using Nvme Raid1 to gain x2 size and resize the DB's to 160GiB. Is it worth to take Nvme faulty? Because I will lose 10HDD at the same time but I have 10 Node and that's only %5 of the EC data . I use m=8 k=2. P.S: There are so many people asking and searching around this. I hope it will work this time. _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx