Calculated value for osd target memory too high for deployments with multiple OSDs per device
#7435
Labels
Comments
janhorstmann
added a commit
to janhorstmann/ceph-ceph-ansible
that referenced
this issue
Mar 15, 2024
The number of OSDs defined by the `lvm_volumes` variable is added to `num_osds` in task `Count number of osds for lvm scenario`. Therefore theses devices must not be counted in task `Set_fact num_osds (add existing osds)`. There are currently three problems with the existing approach: 1. Bluestore DB and WAL devices are counted as OSDs 2. `lvm_volumes` supports a second notation to directly specify logical volumes instead of devices when the `data_vg` key exists. This scenario is not yet accounted for. 3. The `difference` filter used to remove devices from `lvm_volumes` returns a list of **unique** elements, thus not accounting for multiple OSDs on a single device The first problem is solved by filtering the list of logical volumes for devices used as `type` `block`. For the second and third problem lists are created from `lvm_volumes` containing either paths to devices or logical volumes devices. For the second problem the output of `ceph-volume` is simply filtered for `lv_path`s appearing in the list of logical volume devices described above. To solve the third problem the remaining OSDs in the output are compiled into a list of their used devices, which is then filtered for devices appearing in the list of devices from `lvm_volumes`. Closes: ceph#7435 Signed-off-by: Jan Horstmann <[email protected]>
janhorstmann
added a commit
to janhorstmann/ceph-ceph-ansible
that referenced
this issue
Mar 15, 2024
The number of OSDs defined by the `lvm_volumes` variable is added to `num_osds` in task `Count number of osds for lvm scenario`. Therefore theses devices must not be counted in task `Set_fact num_osds (add existing osds)`. There are currently three problems with the existing approach: 1. Bluestore DB and WAL devices are counted as OSDs 2. `lvm_volumes` supports a second notation to directly specify logical volumes instead of devices when the `data_vg` key exists. This scenario is not yet accounted for. 3. The `difference` filter used to remove devices from `lvm_volumes` returns a list of **unique** elements, thus not accounting for multiple OSDs on a single device The first problem is solved by filtering the list of logical volumes for devices used as `type` `block`. For the second and third problem lists are created from `lvm_volumes` containing either paths to devices or logical volumes devices. For the second problem the output of `ceph-volume` is simply filtered for `lv_path`s appearing in the list of logical volume devices described above. To solve the third problem the remaining OSDs in the output are compiled into a list of their used devices, which is then filtered for devices appearing in the list of devices from `lvm_volumes`. Fixes: ceph#7435 Signed-off-by: Jan Horstmann <[email protected]>
|
This issue has been automatically marked as stale because it has not had recent activity. It will be closed in a week if no further activity occurs. Thank you for your contributions. |
Please note the discussion in the linked PR |
janhorstmann
added a commit
to osism/container-image-ceph-ansible
that referenced
this issue
May 25, 2024
Original commit message: Subject: [PATCH] ceph-config: fix calculation of `num_osds` The number of OSDs defined by the `lvm_volumes` variable is added to `num_osds` in task `Count number of osds for lvm scenario`. Therefore theses devices must not be counted in task `Set_fact num_osds (add existing osds)`. There are currently three problems with the existing approach: 1. Bluestore DB and WAL devices are counted as OSDs 2. `lvm_volumes` supports a second notation to directly specify logical volumes instead of devices when the `data_vg` key exists. This scenario is not yet accounted for. 3. The `difference` filter used to remove devices from `lvm_volumes` returns a list of **unique** elements, thus not accounting for multiple OSDs on a single device The first problem is solved by filtering the list of logical volumes for devices used as `type` `block`. For the second and third problem lists are created from `lvm_volumes` containing either paths to devices or logical volumes devices. For the second problem the output of `ceph-volume` is simply filtered for `lv_path`s appearing in the list of logical volume devices described above. To solve the third problem the remaining OSDs in the output are compiled into a list of their used devices, which is then filtered for devices appearing in the list of devices from `lvm_volumes`. Fixes: ceph/ceph-ansible#7435 Signed-off-by: Jan Horstmann <[email protected]>
janhorstmann
added a commit
to osism/container-image-ceph-ansible
that referenced
this issue
May 29, 2024
Original commit message: Subject: [PATCH] ceph-config: fix calculation of `num_osds` The number of OSDs defined by the `lvm_volumes` variable is added to `num_osds` in task `Count number of osds for lvm scenario`. Therefore theses devices must not be counted in task `Set_fact num_osds (add existing osds)`. There are currently three problems with the existing approach: 1. Bluestore DB and WAL devices are counted as OSDs 2. `lvm_volumes` supports a second notation to directly specify logical volumes instead of devices when the `data_vg` key exists. This scenario is not yet accounted for. 3. The `difference` filter used to remove devices from `lvm_volumes` returns a list of **unique** elements, thus not accounting for multiple OSDs on a single device The first problem is solved by filtering the list of logical volumes for devices used as `type` `block`. For the second and third problem lists are created from `lvm_volumes` containing either paths to devices or logical volumes devices. For the second problem the output of `ceph-volume` is simply filtered for `lv_path`s appearing in the list of logical volume devices described above. To solve the third problem the remaining OSDs in the output are compiled into a list of their used devices, which is then filtered for devices appearing in the list of devices from `lvm_volumes`. Fixes: ceph/ceph-ansible#7435 Signed-off-by: Jan Horstmann <[email protected]>
|
@guits Can we merge this? We prepare a back port on our site (osism/container-image-ceph-ansible#488), but I would prefer to merge it upstream. The PR #7502 was closed yesterday because of inactivity. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Bug Report
What happened:
osd target memorywas set to a much higher value after upgrading topacific, resulting in recurring out of memory kills of OSDs.Cause:
Commit 225ae38ee2f74165e7d265817597fe451df3e919 changed the calculation of
num_osds, which is used to calculate a sensible value forosd memory target. The new formula uses ansible'sdifferencefilter, which according to the docs returns a list with unique elements.Thus on deployments with multiple OSDs per device, where the same device should be counted multiple times, the value for
num_osdsis too small and there is an overestimation of the available memory per OSD.Apart from that DB devices are now also counted into
num_osdsWorkarounds:
Set a fixed value for
osd memory targetinceph_conf_overrides.Environment:
ansible-playbook --version):git head or tag or stable branch):stable-6.0(same calculation instable-7.0andmain, but unverified)ceph -v):ceph version 16.2.13 (5378749ba6be3a0868b51803968ee9cde4833a3e) pacific (stable)The text was updated successfully, but these errors were encountered: