The nova-compute fails to start up on a physical hardware if power management is enabled, dedicated cpus are configured, and the nova-compute is restarted (initial start works).
When the strategy is cpu_state then:

• the nova-compute code reads the cpu state and governor values of the dedicated cpus at startup
• then offline all the unallocated dedicated cores (except cpu0)
• then when nova-compute is restarted
• nova tries to read the governor from an offlined cpus and that is not supported by the kernel on a physical core, so nova fails to start up.

https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L853-L859
https://github.com/openstack/nova/blob/master/nova/virt/libvirt/cpu/api.py#L186

physical hardware

root@bedrock:/home/gibi# cat /sys/devices/system/cpu/cpu4/online
1
root@bedrock:/home/gibi# cat /sys/devices/system/cpu/cpu4/cpufreq/scaling_governor
powersave
root@bedrock:/home/gibi# echo 0 > /sys/devices/system/cpu/cpu4/online
root@bedrock:/home/gibi# cat /sys/devices/system/cpu/cpu4/cpufreq/scaling_governor
cat: /sys/devices/system/cpu/cpu4/cpufreq/scaling_governor: Device or resource busy
root@bedrock:/home/gibi# 

We did not hit this when testing in with virtual computes as there scaling_governor file does exists and that is handled in nova: https://github.com/openstack/nova/blob/e82854dc8c514e457528b52834d79176fe5a2135/nova/virt/libvirt/cpu/api.py#L63-L69 But in a physical environment the file exists but kernel returns busy if the cpu is offline.

This is a bug in the power mgmt implementation in nova.