Hi Enrico We were aware of the change on the OSD scheduler before upgrading and changed it back to wpq. We put off the change of this option for another time. However, to make sure this wasn't the root cause of the problem, we set the norecover flag on the cluster, and I suspect that if the recovery process were the reason behind this issue, it would have been mitigated. I think the rocksdb compaction_on_deletion option is what caused the problem, but as, for the time being, we can't put our theory to the test, I'm not quite sure it is what caused it. The reason we suspect so is because I checked the OSD logs and used a Python script that parses the compaction logs and gives useful information about all the compaction operations that happened on the OSD in that log file. I compared the extracted data from logs between two OSDs, one had the option set to true and the other set to false. The number of compactions on the second OSD was almost half of what it was on the first one, and as we had the experience of rocksdb compactions affecting the latency of our OSDs, especially the SSD ones, we came to this conclusion. The compaction data ( https://docs.google.com/document/d/1qXDAHvJnfnOPN4ZevdGbgHmsMse3UpL0gZfSvVic... ) *Do you think this is possible?* And to answer your question about the kind of workload, it is quite diverse. We serve both RBD and S3 clients. Our clients have different workloads, and I cannot describe their behaviour in a certain pattern. This might be a bad idea, and we probably should break the clusters down into several more specific-purpose clusters, but that's a problem for another day. If you have any specific questions in mind that I should answer, *don't hesitate to ask.* And by the increase in the number of compactions on our OSDs, after the option was set to true, we could say that the workload could have high delete requests, which I think is what causes the high number of tombstones. Regards, Nima On Tue, Apr 29, 2025 at 3:40 PM Nima AbolhassanBeigi < nima.abolhassanbeigi@gmail.com> wrote:

Hi Enrico We were aware of the change on the OSD scheduler before upgrading and changed it back to wpq. We put off the change of this option for another time. However, to make sure this wasn't the root cause of the problem, we set the norecover flag on the cluster, and I suspect that if the recovery process were the reason behind this issue, it would have been mitigated.

I think the rocksdb compaction_on_deletion option is what caused the problem, but as, for the time being, we can't put our theory to the test, I'm not quite sure it is what caused it. The reason we suspect so is because I checked the OSD logs and used a Python script that parses the compaction logs and gives useful information about all the compaction operations that happened on the OSD in that log file. I compared the extracted data from logs between two OSDs, one had the option set to true and the other set to false. The number of compactions on the second OSD was almost half of what it was on the first one, and as we had the experience of rocksdb compactions affecting the latency of our OSDs, especially the SSD ones, we came to this conclusion. The compaction data ( https://docs.google.com/document/d/1qXDAHvJnfnOPN4ZevdGbgHmsMse3UpL0gZfSvVic... ) *Do you think this is possible?*

And to answer your question about the kind of workload, it is quite diverse. We serve both RBD and S3 clients. Our clients have different workloads, and I cannot describe their behaviour in a certain pattern. This might be a bad idea, and we probably should break the clusters down into several more specific-purpose clusters, but that's a problem for another day. If you have any specific questions in mind that I should answer, *don't hesitate to ask.* And by the increase in the number of compactions on our OSDs, after the option was set to true, we could say that the workload could have high delete requests, which I think is what causes the high number of tombstones.

Regards, Nima

On Tue, Apr 29, 2025 at 2:33 PM Enrico Bocchi <enrico.bocchi@cern.ch> wrote:

...

Hello Nima,

Unsure if you have found the root cause of the problem in the meantime> From the top of my head, if any useful: - Quincy changes the default scheduler from wpq to mclock - The default number of scrubs on each osd is increased from 1 to 3 - There's a new rocksdb compact_on_deletion option that triggers compaction more frequently according to number of tombstones over a sliding window.

The latter, however, is very workload dependent. What type of workload does the cluster serve?

Cheers, Enrico

...

Hello again We have done some investigating. Since our first message did not have any information or context, let me add some new information.

I looked through our dashboards and the write latency panels in the OSD Overview dashboard. The latency for write operations had increased significantly, but in distinction, the physical write more or less stayed the same; however,

On 4/15/25 12:13, Nima AbolhassanBeigi wrote: the

...

write process operations and write prepare operations have increased dramatically.

In the end, the RocksDB tunings helped our situation and decreased the latency more or less to the pre-upgrade amount. We changed the default parameters to the following:

...

This seems to have done the trick, however, we are still worried that we haven't found the root cause of this problem and only mitigated the symptoms.

Another probably important information is that the bluefs allocator for our OSDs is not the default hybrid and is set to the bitmap option. In addition, the bluestore allocation block score of these OSDs is, for most of them, almost 0.9, which I know is terrible, but what I don't understand is why this problem occurred post-upgrade. Could it be the terrible block scores?

Another interesting change was the memory usage of each OSD

compression=kNoCompression,max_write_buffer_number=128,min_write_buffer_number_to_merge=16,compaction_style=kCompactionStyleLevel,write_buffer_size=8388608,max_background_jobs=4,level0_file_num_compaction_trigger=8,max_bytes_for_level_base=1073741824,max_bytes_for_level_multiplier=8,compaction_readahead_size=2MB,max_total_wal_size=1073741824,writable_file_max_buffer_size=0 post-upgrade,

...

which dropped nearly in half, but the Node's memory cache/buffer usage increased.

I've put the Grafana panels screenshots in this Google Doc < https://docs.google.com/document/d/1DSf4MJoze_BTSWAJWoYyQetA33P6fwpMavYXPdAk...

. We started our upgrade process on March 6 and then again on March 7. We tuned the RocksDB parameters on March 9th, as the screenshots show, which drops the latency.

We are still looking for the root cause, and I could use all the help I can get to try to find it, and it is much appreciated.

Regards Nima

On Mon, Apr 7, 2025 at 3:27 PM Nima AbolhassanBeigi < nima.abolhassanbeigi@gmail.com> wrote:

...

Hi dear ceph community

We have encountered an issue with our ceph cluster after upgrading from v16.2.13 to v17.2.7. The issue is that the write latency on OSDs has increased significantly and doesn't seem to plummet back down. The average write latency has almost doubled, and this has happened since we upgraded the OSDs.

If anybody could help figure this out.

Kind regards Nima

_______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io

-- Enrico Bocchi CERN European Laboratory for Particle Physics IT - Storage & Data Management - General Storage Services Mailbox: G20500 - Office: 31-2-010 1211 Genève 23 Switzerland