Per-CPU stats are more expensive to transport and store, and that
level of detail is not required in many cases.
We export overall total cpu in the same metric as per-cpu, so that
dashboards which previously summed over cpu will work identically.
When cAdvisor starts up, it would read the `vendor` files in
`/sys/bus/pci/devices/*` to see if any NVIDIA devices (vendor ID: 0x10de) are
attached to the node. If no NVIDIA devices are found, this code path would
become dormant for the rest of cAdvisor lifetime. If NVIDIA devices are found,
we would start a goroutine that would check for the presence of NVML by trying
to dynamically load it at regular intervals. We need to do this regular
checking instead of doing it just once because it may happen that cAdvisor is
started before the NVIDIA drivers and NVML are installed. Once the NVML
dynamic loading succeeds, we would use NVML’s query methods to find out how
many devices exist on the node and create a map from their minor numbers to
their handles and cache that map. The goroutine would exit at this point.
If we detected the presence of NVML in the previous step, whenever a new
container is detected by cAdvisor, cAdvisor would read the `devices.list` file
from the container's devices cgroup. The `devices.list` file lists the
major:minor number of all the devices that the container is allowed to access.
If we find any device with major number 195 (which is the major number assigned
to NVIDIA devices), we would cache the list of corresponding minor numbers for
that container.
During every housekeeping operation, in addition to collecting all the existing
metrics, we will use the cached NVIDIA device minor numbers and the map from
minor numbers to device handles to get metrics for GPU devices attached to the
container.
As of the 4.7 kernel, the cpustats field returned from libcontainer
contains values for every possible cpu (including nonexistent ones).
The extra values are all 0s.
If we assume that hotplug events won't happen, we can get a more
accurage cpu count by using runtime.NumCPU and then ignoring any values
beyond that.
If CPU quota is configured (cpu.cfs_quota != -1) the CFS will provide
stats about elapsed periods and throtting in cpu.stats. This change
makes these information available as container_cpu_cfs_* metrics.
Change working set calculation to usage - total_inactive_file, rather than
usage - total_inactive_anon - total_inactive_file. Since writes to tmpfs
get tracked as total_inactive_anon when swap is disabled, the old
calculation would under-report memory pressure.
See this Kubernetes issue for context:
https://github.com/kubernetes/kubernetes/issues/28619
