ceilometer polling agent consumes too much memory on vmware
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
Ceilometer |
Fix Released
|
Undecided
|
Supreeth Shivanand |
Bug Description
ceilometer polling agent compute consumes more than 5gb of memory when configured for a vmware hypervisor.
ps command output:
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
1938 pf9 20 0 5721636 4.391g 1432 S 0.7 57.5 10:26.83 pf9-ceilometerd
Root cause of this issue seems to be because of a change how pollster initializes inspector introduced in pike. Instead of creating only one instance of inspector, it creates one inspector instance per pollster which is a regression and leads to unnecessary memory usage.
In our configuration, we had 50 pollsters in compute namespace and it ended up creating 50 vmware sessions which consumed more than 5 GB memory. I think this issue not specific to vmware and may not be evident on kvm,etc because their session are probably much lighter.
tags: | added: vmware |
Changed in ceilometer: | |
status: | New → In Progress |
assignee: | nobody → Supreeth Shivanand (supreeth90) |
Seeing this since upgrading to Pike release. Probably this change in Pike is related?
https:/ /github. com/openstack/ ceilometer/ commit/ 66179f89a065116 8d59cd9c36bac5f 9ffb104471