Clarify hypervisor CPU metrics
| Affects | Status | Importance | Assigned to | Milestone | |
|---|---|---|---|---|---|
| nova-dpm |
Triaged
|
Medium
|
Sreeram Vancheeswaran | ||
Bug Description
DPM is using shared IFL processors. DPM was configured to use 2 processors:
nova hypervisor-show foo
+------
| Property | Value |
+------
| cpu_info_arch | s390x |
| cpu_info_features | [] |
| cpu_info_model | host |
| cpu_info_
| cpu_info_
| cpu_info_
| cpu_info_vendor | IBM |
| vcpus | 2 |
| vcpus_used | 0 |
+------
Now if you start an instance with 2 processors the output looks like this:
nova hypervisor-show foo
+------
| Property | Value |
+------
| cpu_info_arch | s390x |
| cpu_info_features | [] |
| cpu_info_model | host |
| cpu_info_
| cpu_info_
| cpu_info_
| cpu_info_vendor | IBM |
| vcpus | 2 |
| vcpus_used | 2 |
+------
If you start another instance with 2 processors the output still looks like this:
nova hypervisor-show foo
+------
| Property | Value |
+------
| cpu_info_arch | s390x |
| cpu_info_features | [] |
| cpu_info_model | host |
| cpu_info_
| cpu_info_
| cpu_info_
| cpu_info_vendor | IBM |
| vcpus | 2 |
| vcpus_used | 2 |
+------
--> That's confusing! Should we reconsider the the way of how vcpus are calculated?
The issues probably is, that we have shared CPUs, but each partition can only have be attached to a single shared CPU once. In other words, if you have 3 shared cpus, a partition can have a maximum number of 3 vCPUs.
In addition the other CPU attributes should be clarified as well! (cores, sockets, threads,...)
| description: | updated |
| Changed in nova-dpm: | |
| assignee: | nobody → Sreeram Vancheeswaran (sreeram-vancheeswaran) |
| importance: | Undecided → Medium |
| status: | New → Triaged |
| summary: |
- Number available and used vcpus is confusing + Clarify hypervisor CPU metrics |
| description: | updated |
| Marco Pavone (pavone) wrote | #1 |
I agree it is confusing, but this seems like a general OpenStack issue: https:/
The most interesting question is, if that is "only" a UI/API deficiency, or if the scheduler is relying on wrong data, i.e. if the pCPU * Overcommit-Ratio is not the same as total vCPUs - so the impact would be that not all avail. resources could be used.