Hi everyone, This is my first time posting here, so it's nice to meet you all! I have a Ceph cluster that was recently upgraded from Octopus to Pacific and now the write performance is noticeably worse. We have an application which continually writes 270MB/s through the RGW which was working fine before the upgrade. Now it struggles to write 170MB/s continuously. I don't have detailed benchmarks from before the upgrade other that what I just mentioned but I have investigated some since. For background the cluster is two hosts with 10 HDD OSDs each which is under 5% usage overall. The RGW pool is set up with a k=4, m=2 erasure code (across OSDs). Profiling with `rados bench` to a pool with the same erasure code setup gives similar 170MB/s performance (so about 42MB/s per disk). Profiling a single OSD using `ceph tell osd.0 bench` yields an average of 40MB/s across all disks (which should yield close to 160MB/s at the pool level). Given this it seems to me to be an issue at the OSD level. I tried setting `bluefs_buffered_io` to false as mentioned elsewhere. This reduced the `%wrqm` reported by iostat (which was previously close to 100%) and gave a slight performance gain (around 50MB/s per OSD), but nothing close to what was seen previously. Before and after the above change both iostat and iotop report over 100MB/s written to disk while a single `ceph tell osd bench` command is running and iotop shows a near threefold write amplification. With the benchmark writing 1GB to disk iotop shows both the `bstore_kv_sync` and `bstore_kv_final` threads writing about 1GB each and the threads `rocksdb:high0` and `rocksdb:low0` writing a total of 1GB. So 3GB total for the 1GB benchmark. Looking at the OSD logs during the benchmark it seems rocksdb is compacting several times. I tried adding sharding to one of the OSDs as mentioned in the documentation (with `ceph-bluestore-tool`) but it didn't seem to make a difference. Does anyone have any idea what may have caused this performance loss? I am happy to post any more logs/detail if it would help. Thanks! Dustin _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |